Digital Pixel-Accurate Intensity Processing Method for Image Information Enhancement

This application is related to application Ser. No. 08/207488 filed contemporaneously herewith, entitled "Two Dimensional Digital Hysteresis Filter for `Smoothing` Digital Images" invented by Klaus-Ruediger Peters, Eisaku Oho, William Martin and David R. Jenkins (Attorney Docket No. 93-1365).

This invention relates generally to the processing of digital image data. More specifically, this invention relates to image data processing using digital techniques for image enhancement and enlargement.

Digital image processing has taken an increasing importance as a result of the technological advances in image acquisition and image communication; and can provide advantages over conventional analog image information handling, e.g., undisturbed access of the "raw data set", objective image evaluation, quantitative analysis of the image information, and reduced costs and increased flexibility of image data handling. While a complex infrastructure network is in place for high-speed image communication with local, national and international access, the general use of digital image processing is hindered through lack of universal standards for identifying image information. Unfortunately, visual image perception varies amongst individuals and depends strongly on the image perception and pattern recognition ability. This is the reason why the same image is evaluated quite differently by more than one person. Of course, the lack of consistency in image information analysis and display is extremely problematic and creates serious concerns for image evaluation.

There are many kinds of information contained in images, but only a few classes may be important in image communication, i.e., detail-oriented (scientific, technical and medical images), composition-oriented (arts, materials science) or information-oriented (binary and CAD, computer assisted drawing). Most important are detail containing images which describe two-or three-dimensional data sets dealing with spacial features. The lack of proper pixel accurate tools for objective description of image details as well as image imperfections produced by acquisition and transmission (noise) limit image communication at this time to information-oriented contents only.

When analyzing images (e.g., microscopic images derived from SEM or TEM; or medical imagery such as mammograms or x-rays), the image analysts are limited most by their own visual system (e.g., the human eye) regarding image perception and pattern recognition, since most modern imaging instruments (e.g., microscope, x-ray device, mammography device) provide more data than the eye can process. In general, such data are spacial information documented with certain contrast mechanisms and translated into images. Irrespective of the kind of imaging technique used, the imaging information must be communicated to the visual system for evaluation of its information content at the level of imaging instrument contrast resolution as well as spacial resolution. Since the visual perception is limited in intensity range (IR) to 100-200 intensity levels (7-8 bit in self illuminated images of a video monitor) and resolution to 500-1,000 pixels per picture width (PW), the imaging instruments full frame image information must often be compressed to fit within these values. Image evaluation is primarily a process of pattern recognition which works at a much lower information density than the eye can perceive. Only patterns of large image components of high contrast, high edge sharpness and a few intensity levels (4 bit) are recognized. All other small detail information can only be partially recognized and therefore is commonly generalized as image background or texture. However, in microscopy, radiology, x-ray and other imaging sciences, such background contains a wealth of information of acquired image detail data which is commonly lost in visual analog evaluation. Simply increasing the image magnification during acquisition will not fully access these details due to visual or instrumental limitations (low contrasts, sample or instrument instabilities, etc.). It is the ability of digital image processing to make these details visually accessible (detail enhancement) in already acquired images and to provide simple tools for their evaluation, quantitation and communication which makes such processing an important and essential tool for image evaluation. Unfortunately, as will be discussed below, presently available and known digital processing techniques fall far short of providing the required detail enhancements and are associated with serious drawbacks and deficiencies.

In single digital images, intensity fluctuations caused by random noise and small structural (spacial) image details of only a few pixels cannot be identified since no information other than the pixel's intensities are available in a single data set. However, digital image processing methods can be applied for separating image contents on account of certain intensity criteria. The separation of noise and detail structures can be defined by a threshold of intensity variations below which spacial variations are thought to represent noise and are eliminated (smoothed) but above which the intensity variations are defined as significant and are maintained. Conventional image processing methods apply spatially extended processing masks or Fourier filters for the determination of local intensity fluctuations (either in the space domain or in the Fourier domain) and use various methods for determination of the threshold intensity value. The utilized pixel area (mask) and an often used weighing factor applied to the local intensity distribution within the mask will shift the spacial boundary between (smoothed) background and (maintained) detail, altering the spacial dimensions of details. Using this technique, serious spacial artifacts are produced when structural features are similar in size or smaller than the effective spacial filter area. This problem occurs in all conventional processing modes using spacial kernels, Fourier filters or statistical approaches. In addition, certain significant spacial intensity distributions (significant structural patterns) may be seriously altered by eliminating or adding structures, indicating a strong dependency of the processing result on the image content. Such alteration of the spacial content of the original image (raw data set) is a serious limitation of all conventional noise filters in cases where the structural integrity of the image data is important--i.e., in structure characterization and quantitation. The extent of spacial artifacts in G7 conventional processing depends on the image content. Therefore, complex time-consuming determination of optimal processing parameters are required for each image in order to reduce processing artifacts. Also, conventional image processing speeds are so slow that visual control of intensity threshold adjustments and recognition of processing artifacts are significantly hindered. Therefore, conventional noise smoothing techniques are applied strictly to full frame images, thereby reducing the visualization and recognition of the produced artifacts.

The above-discussed and other problems and deficiencies of the prior art are overcome or alleviated by the digital imaging process method and system of the present invention. In accordance with the present invention, novel "smoothing" and "detail enhancement" techniques for processing digital images create a new quality of image perception (centering on enhancement of objective image information) which have wide application in the general field of image enhancement and more particularly in the fields of microscopy, radiology, remote sensing, astronomy, robotics and machine vision and general image communication. The new image processing technology of this invention makes possible a separation of the image information into three objective distinct non-random intensity groups of "large image features" and small "spacial detail" and "intensity detail", and one group of random intensity information of the "image noise". "Smoothing" allows an elimination of intensity-defined image components and the concomitant reduction of image contrast can be restored by linear contrast stretching of the smoothed image. "Detail enhancement" allows a recovery of intensity-defined image components eliminated by the smoothing process through subtraction of the smoothed image from the original image. The overall contrast range of the recovered information is equal to the applied smoothing factor and can be automatically brought to the full visual intensity range by linear contrast stretching. This has the advantage of maintaining detail contrast proportions which closely reflect the raw data characteristics. If high pixel accuracy processing is provided, "detail slicing" becomes possible. Two smoothed images, each processed with a different smoothing factor, can be subtracted providing the information difference between least smoothed and the most smoothed data set. Again, precise linear contrast stretching can present the extracted information at full visual intensity range.

The smoothing technique of this invention comprises a two dimensional digital hysteresis filter which utilizes a variable, automatically adjusting "two-dimensional mask". As a result, the filter is independent of the image size and content and cannot alter the size of any structural (significant intensity) features, even if they are as small as only one pixel in size. In one embodiment of this technique, the two-dimensional hysteresis smoothing technique of this invention calculates smoothed pixels using a set of one-dimensional hysteresis lines at various angles running through each pixel in an image. Each one-dimensional hysteresis line calculates a new value for each pixel in an image. These one-dimensional hysteresis values are added together and divided by the number of hysteresis lines to get the output value for the pixel. This operation requires complex processing and is preferably realized through a massively parallel processing array processor which allows near-real time full processing speeds on 1K.times.1K images which may require several billions of instructions per record (BIPS). The technique of this invention is preferably implemented on a desktop massively parallel processor that has a large number of 16-bit processing elements (PE's) connected in a ring. Each PE has a small, high-speed on-chip data memory and a large off-chip data memory. The PE's all have access to a shared scaler data memory and a shared program memory. A single program is executed by all the PE's in lockstep (SIMD processing).

As used herein, "smoothing" thus constitutes a method for reducing or smoothing selected intensities in a digitized image data array comprising a matrix of pixels arranged in columns and rows, including the steps of:

(a) determining if a selected pixel's intensity value in relation to its neighbors' intensities falls within a preselected significant intensity range or cursor width which equals a fixed two-dimensional mask one pixel in size;

(b) maintaining the selected pixel's relative intensity value if it is outside the cursor width;

(c) changing the selected pixel's intensity value if it is within the cursor width to an intermediate value; and

(d) repeating steps (a)-(c) for at least some of the other pixels in at least a portion of the matrix to define a smoothed digitized image data array.

A somewhat different smoothing technique in accordance with this invention constitutes a method for reducing or smoothing selected intensities in a digitized image data array comprising a matrix of pixels arranged in columns and rows including the steps of:

(a) determining if a selected pixel's intensity value in relation to its neighbors' intensities falls within a preselected significant intensity range or cursor width;

(b) maintaining the selected pixel's relative intensity value if it is outside the cursor width;

(c) changing the selected pixel's intensity value if it is within the cursor width to an intermediate value which is an average of the selected pixel's intensity value relative to its neighbors' intensity values, the average being determined by using a set of one dimensional hysteresis lines;

(d) determining the area of the pixel matrix which contributes to the final intermediate value of a pixel. This area may vary for each pixel and is defined solely by the image's intensity characteristics. This area represents a two-dimensional variable mask which is generated automatically by the one-dimensional mask but which may not be identical with the finally smoothed area of the pixel matrix; and

(e) repeating steps (a)-(d) for at least some of the other pixels in at least a portion of the matrix to define a smoothed digitized image data array.

Preferably, the hysteresis lines comprise spatially neighboring pixels. More preferably, the hysteresis lines are linear lines radiating at preselected angles through said selected pixel.

The smoothing technique of this invention which utilizes a "two-dimensional hysteresis filter" provides many features and advantages relative to conventional digital imaging techniques. For example, conventional image enhancement using fixed small processing masks (kernels in space domain or Fourier filters) may not be pixel accurate (as is the technique of the present invention). Spacial artifacts may be as large as the size of the mask and depend strongly on the image content. In order to reduce such artifacts, the processing parameters must be optimized for each image. However, if the image itself is used for the definition of a local processing mask (as in the present invention) such spacial artifacts can be eliminated and the processing result will be independent from the image content. As in conventional mask processing, the processed pixel must be related to all other pixels within the "mask" assessing the spacial significance of intensity differences by the smoothing factor. The use of a "variable automatically adjusting local mask" produces a processing characteristic equal to point processing since each processed point has a specifically adjusted spacial mask. High precision processing reduces processing artifacts to a level not visible in enhanced images.

In general, "detail enhancement" refers to an enhancement of the contrast of image details and must include the spacial details as well as the intensity details in order to maintain the image character (image accuracy). "Spacial details" constitute intensity variations over a short distance (a few pixels long), and "intensity details" constitute intensity variations of a few intensity steps independent of their spacial extent. In addition, a desirable enhancement procedure must preserve the unrestricted possibility of image quantitation not only of the spacial content but also of the intensity content; that means the image processing technology must maintain the accuracy of the image at the level of individual pixel's intensity. Such pixel accurate intensity processing (PAIP) for image enhancement is not possible with any conventional technology, but is achievable utilizing the detail enhancement processing technique of the present invention which can fulfill the latter requirements and which is therefore inherently suitable (trustworthy) for scientific and medical applications. The basis for its spacial accuracy is the utilization of pixel-accurate intensity processing; which preferably utilize the "smoothing" technique discussed above. Its processing accuracy results from the capability of the "enhancement parameter" (significant intensity range) of selecting "intensity information" from the whole image at both levels of spacial and intensity details. Its application power results from the possible visual appreciation of the spacial and intensity changes of selected information by linear contrast stretching to the full intensity range of human visual perception, and from the possibility of performing the enhancement in "near-real time" which provides the selected information at a time interval similar or equal to the processing speed of human visual recognition taking full advantage of the visual capabilities for image evaluation and communication.

The novel PAIP detail enhancement filter of the present invention works in the following way. First, the detail content of the image is reduced on the basis of intensity variations defined by a single processing parameter consisting of the "cursor width" using the two dimensional hysteresis smoothing technique described above (or any other "pixel-accurate" smoothing technique). In this application, significantly larger cursor widths are used than in conventional noise management since the image content must not be maintained but instead must be reduced. Secondly, the detail image is created by subtracting the smoothed image from the original image (or from another smoothed image). The maximum intensity difference of the detail image between any points is equal to the applied cursor width (or the difference of the larger minus the smaller cursor width). Thirdly, the contrast range of the detail image is enhanced by linear contrast stretching with maximum enhancement obtained by utilizing the full width of the intensity range available for visual perception (typically 8 bit). The evaluation of the processing result is dramatically enhanced if the processing occurs in "near-real time" (less than one second) so that an interactive change of the processing parameter becomes possible. This enhances the perception of the image details and their correlation with the original image.

As used herein, "detail enhancement" thus constitutes a method of enhancing the detail in a digitized image data array comprising a matrix of pixels arranged in columns and rows including the steps of:

(a) generating two pixel-accurate image data arrays with at least one of the image data arrays being smoothed; and

(b) subtracting one of the pixel accurate arrays from the other pixel accurate array to define a differential image having a selected intensity range.

An important and often (though not always) critical third step includes:

(c) enhancing the contrast of the differential image through contrast stretching.

Specific image detail recognition is a new visual task for image evaluation and communication which conventionally requires long visual experience. The "near-real time" processing fosters a quick and realistic training of the visual system (e.g., the human eye) for working with the complete image data set. Artifact-free processing is a prerequisite for such an expansion of the visual perception not achievable by conventional image processing techniques.

The digital enhancement technique of the present invention provides many features and advantages relative to conventional digital imaging techniques. For example, the detail enhancement filter works principally different from other conventionally used detail filters, i.e., the Oho filter which is described in Oho E., Automatic Contract Adjustment for Detail recognition in SEM Images On-Line Digital Image Processing, Scanning 14: 335-334 (1992). The Oho filter is a highlight filter for edge enhancement of spacial image details only (it specifically suppresses intensity detail enhancement). It extracts small area intensity changes (highlights) from the original image by use of a fixed large mask size median filter (19.times.19 for 1K.times.1K images). The use of a median filter has many limitations. Most importantly are the following three aspects; first, the filter maintains local intensity variations and when the median is subtracted from the original image, the intensity details are completely removed from the selected detail data. Second, the filter cannot select a certain intensity range. This has the serious implication that the intensity range of the detail image depends fully on the image content. Therefore, only histogram equalization can be used for contrast enhancement of the filtered details which may cause spacial contrast artifacts. Third, the Oho filter erodes fine structures at the level of its mask size at edges and at small multiple intensity variations. As a consequence, the enhanced detail image must be added to the median filtered image in order to restore (in part) the eroded fine structure. However, since the detail image contrast is enhanced through disproportional histogram equalization, not all spacial details are enhanced proportionally to their original intensities. In contrast, the new PAIP detail enhancement filter of this invention produces quite different information as compared to that obtained by the Oho filter and, presents all image information in an exhaustive fashion with linear contrast enhancement.

The new detail filter produces detail images which characterize the intensity distributions within an image. It provides a tool for a novel method of image information analysis and classification based on the concept that any image communicates information only through image contrasts which are intensity variations between certain pixel arrays within the total data matrix. An area of certain contrast is defined by the differences between its average intensity and the surrounding intensity irrespectively of the overall intensity variations (background). The new detail filter of this invention provides the only tool available for selecting such local intensity variations independently from the other intensity variations within the full data matrix through only one parameter which is the intensity range (equal to cursor width or significant intensity range). Application of the filter with increasing intensity ranges selects from the image contrast components of specific visual information contents in a defined and principle manner which matches the visual pattern recognition mechanisms. In general, digital images represent two- (or three-) dimensional intensity maps which characterize the spacial x/y(/z) location of all contrast information. Successive smoothing of the intensity map with increasing intensity ranges and extraction and contrast enhancement of the intensity variations between the smoothed maps produces contrast patterns which have dramatically different visual impact so as to define specific different and characteristic information classes. Only three different non-random intensity variations are evident in any image, which characterize (1) large spacial image components (features) of high contrast and (2) small variations either of spacial (spacial details) or (3) intensity character (intensity details). In addition, smaller random intensity variations are perceived (noise). Visually not perceivable is a constant background intensity level (background) which has no spacial information due to the lack of contrast. The new definition of these discrete image intensity components is important since they are not spatially limited to certain image contents but can be used to characterize all spacial information of any image with highest spacial accuracy (including one pixel wide features). The intensity classes defined in this way match the visual information recognition principles and are therefore uniquely suited for the information analysis of images generated by scientific, microscopical, medical and industrial technologies. These technologies communicate specific information through a variety of contrast mechanisms, the origin of which is often non-visual and purely scientific in nature (phase shifts, diffraction and adsorption, electron spin modifications and any other way of distinguishing spatially related features from each other). However, when the information contents are displayed as visual images they must be translated into an intensity map recognizable by the visual system. Since the described detail filter of the invention generates all basic visual information classes it is best suited for information analysis, quantitation and communication of such technical images.

The ability of categorizing and quantifying image information also provides a tool for objective measurement of image quality. For this purpose, the intensity ranges (significant intensity ranges) of each image intensity class are proportionally distributed within the overall intensity range of the raw data in the sequence as generated and displayed by the detail filter: first the noise, then the spacial detail, the intensity detail, and the image feature/background. A graphic display of the image information components and their proportion within the intensity range as schematic intensity maps in an "information cube" facilitates visual image quality assessment (see FIG. 16). Either the proportion of the noise component to the adjacent intensity information of any width, or the relation of any component to any other or the overall intensity range provides an objective tool for categorizing images and image quality. Such graphic display of the image information content will facilitate image analysis and communication.

Based on the foregoing, it will be appreciated that important features of the smoothing techniques of this invention include:

(1) preserves the original dimensions (measured in pixels) of smallest details with an accuracy of individual pixels (pixel-accurate);

(2) is independent of the image content and image size;

(3) uses only one simple numerical parameter which describes the intensity threshold for the filter;

(4) minimizes processing artifacts; and

(5) can be performed at "near-real time" in order to overcome the limitations of conventional image evaluation.

Similarly, important features of the digital intensity enhancement technique of this invention include:

(1) image accurate (addresses spacial and intensity details);

(2) pixel accurate (does not change spacial dimensions at single pixel level, maintains the linear relation of image intensities, required for quantitation and pattern recognition); allows pixel accurate intensity processing (PAIP);

(3) independence from the image content (applies to all images); and

(4) one "processing parameter" (significant intensity range) with easy adjustment by visual control through "near-real time" processing.

(5) objective definition from the characteristics of intensity variations of image information classes which include a) image features, b) spacial details, c) intensity details and a random class of image noise.

(6) objective quantitation of image information and image quality using the proportion of each information class with respect to the full image intensity range.

(7) graphic display of image characteristic with an "Info-cube" which presents a schematic view of the extent and proportion within the image intensity range of the information classes present in the image.

The above discussed and other features and advantages of the present invention will be appreciated and understood by those skilled in the art from the following detailed description and drawings.

Referring now to the drawings, wherein like elements are numbered alike in the several FIGURES:

FIG. 1A is a block diagram illustrating the process of the present invention;

FIG. 1B is a block diagram of a system in accordance with the present invention;

FIG. 1C is a flow chart depicting the smoothing technique of the present invention which utilizes a two-dimensional hysteresis filter for noise reduction;

FIG. 2 is a diagrammatic example of 45 degree hysteresis lines covering an image;

FIG. 3 is a flow chart of the "group" processing operation used in the smoothing technique of the present invention;

FIG. 4 is a diagrammatic representation of the data structures used in the smoothing technique of the present invention;

FIG. 5 is a graph depicting the relation of digital image information with visual pattern recognition parameters;

FIG. 6 depicts in Section "A" the principle approach for intensity information extraction; and in Section "B" a group of graphs depicting different types of intensity defined information including "noise", "spacial detail", "intensity detail" and "feature/background" which are contained in digital images and in Section "C" a schematic representation of the intensity extent of the various information groups;

FIGS. 7A-F depicts a test pattern of Gaussian noise having been processed using the noise smoothing technique of this invention with and without prior randomization of the image;

FIGS. 7G-H depicts a test pattern of Gaussian noise and conventional noise reduction techniques;

FIGS. 8A-H are images of low magnification SEM data depicting the information classes and pixel-accuracy of the detail enhancement using the detail enhancement techniques of this invention;

FIGS. 9A-B are images of high magnification SEM data having been processed using the detail enhancement techniques of this invention;

FIGS. 10A-B are images of high magnification field emission SEM data having been processed using the detail enhancement techniques of this invention;

FIGS. 11A-B are low voltage field emission SEM data having been processed using the detail enhancement techniques of this invention;

FIGS. 12A-B are environmental SEM data having been processed using the detail enhancement techniques of this invention;

FIGS. 13A-B are high resolution field emission SEM data having been processed using the detail enhancement techniques of this invention;

FIGS. 14A-D are cryo-TEM energy filtered, phase contrast data having been processed using the detail enhancement techniques of this invention;

FIGS. 15A-B are high voltage dark field TEM data having been processed using the detail enhancement techniques of this invention;

FIGS. 16A-B are high resolution TEM data having been processed using the detail enhancement techniques of this invention;

FIGS. 17A-B are high magnification scanning transmission electron microscopy data having been processed using the detail enhancement techniques of this invention;

FIGS. 18A-B are images of atomic force microscope or AFM data having been processed using the noise smoothing and detail enhancement techniques of this invention;

FIGS. 19A-B are images of light microscope data having been processed using detail enhancement techniques of this invention;

FIGS. 20A-D are images of confocal laser light microscopy data having been processed using the noise smoothing and detail enhancement techniques of this invention;

FIGS. 21A-D are images of mammogram data having been processed using the noise smoothing and detail enhancement techniques of this invention;

FIG. 22A-D are images of chest x-ray data having been processed using the noise smoothing and detail enhancement techniques of this invention; and

FIG. 23 is a group of graphs depicting quantitation of image data "information cubes" using the smoothing and detail enhancement techniques of this invention.

The present invention comprises several related digital image processing techniques including a novel "smoothing" or "data reduction" technique which utilizes a two-dimensional hysteresis filter for noise reduction and a novel "intensity enhancement" technique which enhances the "smoothed" or "reduced" data for selected spacial details and intensity levels. The novel "smoothing" technique (which may also be used in the "detail enhancement" technique) will be described first.

A. Prior Art One-Dimensional Hysteresis

The use of a "smallest one-dimensional mask" was suggested in the techniques of a "linear hysteresis smoothing procedure" Ehrich, R. W. (1978), A Symmetric Hysteresis Smoothing Algorithm That Preserves Principal Features, Computer Graphics and Image Processing 8, 121-126). These techniques were developed for linear (one dimensional) spectral data, i.e., I/x (I=intensity at the point x along a x-direction). This prior art linear procedure uses a cursor of certain intensity height (cursor "width") which is moved along the data set. A fixed reference point on the cursor (normally its center) is used for determining the resulting intensity--i.e., the new intensity value at a specific point in x direction. The original data intensities are compared to the maximum and minimum value of the cursor and determined to be either outside or equal to/within the cursor width. In the first case, the cursor position changes and follows with one of its endpoints the data until the intensifies are again within the cursor width. In the latter, the cursor intensity position is not altered until data values outside of the cursor values are again encountered, contributing to the smoothing effect.

These one-dimensional procedures produce a linear new data set which has reduced contrast with an overall intensity reduction by as much as one half of the cursor width (between smoothed areas and remaining data). However, it contains all of the sequences of the original data set where the data values were outside the cursor width. In between these regions, a new and constant value replaces regions where the intensity fluctuations were within the cursor width. These smoothed regions have an unchanged constant intensity of a value that has no relation to that of the previous or following smoothed regions and that is only dependent on the last intensity maintained from the raw data set. This is the reason why the procedure generates linear "streaks" of various intensity values. Behind high intensity features (top of cursor) a "top streak" is produced as well as a "bottom streak" at the level of low intensity variations (bottom of cursor). The length of these streaks depends upon the linear characteristics of the data set. In order to recover raw data which are obscured at the beginning of streaks (at the top as well as at bottom streaks), the linear (e.g., one-dimensional) smoothing technique of the prior art is performed in reversed direction, and both new data sets are arithmetically averaged maintaining their cross-registration.

The dual direction processing of the prior art has significant advantages and disadvantages. Since the maximum intensities of a structure ("top" and "bottom") are maintained when read in both directions, the height and position of the structure is maintained in the averaged data (only reduced or increased from the raw intensity by the value of the reference point on the cursor). At both shoulders of the structure, within the streaks, the slope of the raw data set is modified along the streak until again the raw data are read providing for the continuity of the smoothed data set. In some regions, the streak intensity may be different in both directions and an averaged "background" intensity will result. In order to minimize border artifacts at the beginning and end of the linear data set, at each start of reading, the cursor reference point is positioned on the original data value. If this procedure would put the cursor reference point outside of the total intensity range, then the cursor is positioned so that it stays at the maximum or minimum intensity value of the raw data. The disadvantage of this procedure is that the background streaks have no common reference and are strongly identifiable by their linearity independent of their intensity value. For one dimensional data, this is not a problem, since the visual system can easily interpolate all flat regions into a common varying background assessing its overall long frequency behavior.

Application of this one-dimensional smoothing technique to images have been attempted by reading and processing each line of an image data set independently and replacing the raw data by the averaged resulting intensity values. This procedure has been associated with disadvantages and problems and has produced serious image artifacts (Duda and Hart, 1973) which are not reduced even if the two-dimensional data set is read as one dimensional continuum in zig-zag fashion. The resultant image is full of large and wide streaks which totally disrupt the visual perception of the two-dimensional data.

B. The Present Invention--Two-Dimensional Hysteresis

The present invention utilizes a novel digital processing technique to fully reduce and minimize these "streaking artifacts" of the one dimensional processing of image data through the development of an automated "two-dimensional" processing routine which maintains all advantages of the one dimensional hysteresis approach and expands their utilization to two-dimensional data--i.e., I/x,y; I=Intensity (or any other) values, x-, y-values=coordinates in x/y plane.

Since the intensity value of the top and bottom streaks depend on the last low intensity point of the raw data, and since this value can randomly fluctuate from noise, many different raw data end points must be generated and then averaged. This is not possible in one-dimensional data sets because of the lack of additional correlation of the data other than its one directional sequence. However, in two dimensional data, many spacial correlations can be generated within the two dimensionally defined image areas. One such correlation mechanism can be the direction in which the image data is read--i.e., the angle between the reading and the data axis. Other correlations may be found in spiral reading frames or alternating forward and backward directions or by any other way within the two dimensional image, which will address all image points and maintain their spacial correlation. Reduction or lack of sequential spacial correlation will weaken the smoothing effect of the procedure. The processing of an image of varying background levels after randomization and the derandomization of the processing product produces no smoothing whatsoever, but only a contrast reduction by a value equal to that of the effective cursor width demonstrating the requirement of spacial continuity of the data and the high sensitivity for the image content.

C. Implementation of Two-Dimensional Hysteresis Smoothing Technique

In a preferred embodiment, the two-dimensional hysteresis smoothing technique of this invention calculates smoothed pixels using a set of one-dimensional hysteresis lines at various angles running through each pixel in an image. The technique requires an input image and two parameters, (1) the number of hysteresis lines per pixel, and (2) the cursor width for the hysteresis algorithm. The program calculates an output image of the same size as the input image. In an actual implementation of this invention, a suitable computer program was written in assembly language for the AP.sub.x desktop massively parallel processor manufactured by Visionary Systems Inc. of New Haven, Conn. and described in detail in the paper entitled "The AP.sub.x Accelerator", E. Abreu et al, Proceedings of the 2nd Symposium on the frontiers of Massively Parallel Computation, Fairfax, Va., Oct. 10-12, 1988 IEEE Computer Society Press. It is called from a C language subroutine which runs on a PC-AT clone. The source code for both the PC-AT and the AP.sub.x is included in Table 1 below. The AP.sub.x parallel processor has up to 256 16-bit processing elements (PE's) interconnected in a mesh topology. Each PE has a fast 256 word on-chip data memory and a 65536 word off-chip data memory. The PE's all have access to a shared scaler data memory and a shared program memory. A single program is executed by all the PE's in lockstep (SIMD processing). The AP.sub.x uses a PC-AT clone as a Host computer, and the AP.sub.x off-chip PE memories are memory-mapped into the Host computer's address space.

A flowchart of the overall two-dimensional hysteresis smoothing program of this invention is shown in FIG. 1C. It performs eight passes over an input image, with each pass corresponding to hysteresis lines in a particular angle range. Table 2 shows the characteristics of each angle range. The actual hysteresis lines are at equally spaced angles around a 360 degree circle. For example, if the number of hysteresis lines is 16, there will be hysteresis lines at angles of 0, 22.5, 45, 67.5, 90, 112.5, . . . , 270 and 292.5 degrees and they will be partitioned into eight angle groups of two lines each. FIG. 2 shows an example of 45 degree hysteresis lines covering an image. Each hysteresis line calculates an output value for each pixel in the image. The final output value for a pixel is the average of all the output values for that pixel.

The program does one pass over the input image for each of the eight angle groups, doing the calculations for all the hysteresis lines within that group. A flowchart of the group processing is shown in FIG. 3. The program steps sequentially through the image rows beginning with the Starting Edge from Table 2. If the Starting Edge is either Left or Right, the input and output images are transposed before and after the pass, so that the program can step from row to row by incrementing or decrementing its memory address. Each row is read in turn and each hysteresis line in the group is applied to it. Hysteresis midpoints are then conditionally shifted to the left or fight neighboring PE's, to maintain the correct angle of the hysteresis line, and the process repeats for the next line. The conditional shifting is based on the line patterns stored in the scaler memory.

The hysteresis calculation is as follows: if the input pixel value is less than the current cursor midpoint minus half the cursor width, the midpoint is changed to the input pixel value plus half the cursor width; if the input pixel value is greater than the current cursor midpoint plus half the cursor width, the midpoint is changed to the input pixel value minus half the cursor width. The output value for the pixel is the resulting cursor midpoint. This is added to the midpoints for other hysteresis lines applied to the same pixel. At the Starting Edge, the cursor midpoints are initialized to the input pixel values. When cursor midpoints are shifted off the edge of the image, they are re-initialized with the value of the edge input pixel. Initialized cursor midpoints are clamped to their valid range; from the minimum possible pixel value plus half the cursor width to the maximum possible pixel value minus half the cursor width.

The input and output images are stored in the large, off-chip PE data memories. Each PE stores at least one column of the image in its memory as shown in FIG. 4. In this description, it is assumed that the number of PE's is the same as the number of columns in the image to be processed and that each PE stores one column from each row. If there are more PE's than columns, the extra PE's are simply disabled and do not participate in the calculations. If there are more columns than PE's, each PE does the processing and storage for multiple columns. During processing, each output pixel holds a sum of the values calculated by the hysteresis lines running through the pixel. The final output image is calculated by dividing the pixel sums by the number of hysteresis lines.

Each on-chip PE data memory holds cursor midpoints for all the hysteresis lines running through the current pixel in the current angle group. It also temporarily stores each input row as it is being processed, and the output sums for the pixels in that row.

The scaler data memory holds line patterns for sets of eight hysteresis lines. One line pattern represents eight hysteresis lines whose angles are offset in increments of 45 degrees. This pattern represents the angle of the hysteresis line in terms of vertical/horizontal and diagonal steps of one grid unit. It is used to control the conditional interprocessor shifting of cursor midpoints in the hysteresis processing.

The optimal number of iterations (of new reading frames necessary for artifact suppression) will depend on the raw data set and/or the cursor width. Insufficient averaging will produce visually destructive contrast jumps along the reading directions. Such artifacts are easily seen in straight (radial) reading directions and in images of non linear structural characteristics. Eight to thirty two different linear reading directions may be sufficient in reducing the artificial linear background intensity fluctuations so as to be non-recognizable by contrast analysis in average images even if small features of high contrasts are present. In addition, for the rare cases where streaks are persistent, sixty four to two hundred fifty six (or more) iterations are provided. The number of iterations required can be set automatically from the chosen cursor width according to a semi-empirical evaluation of effectiveness or from calculations using the selected cursor width and contrast properties of the image. However, other non-linear reading directions may be provided in order to distribute the background intensity fluctuations in such a manner that it will not be recognized by visual perception. Such a non-linear method may be applied to an unsatisfactory result of the linear method for the purpose of redistribution of the background intensity fluctuations into a non recognizable or non distracting pattern.

A. Information Content Of a Digital Full Frame Image

Referring to FIG. 5, the information content of a digital full frame image can be categorized into a non-random (structural) class of features and details, and a random class of noise. A feature is an image component accessible to visual pattern recognition, i.e., it is larger than .about.10% of the picture width, has a contrast range of more than .about.5% of the visual intensity range, and each must fall within no more than 15-20 intensity levels in order to be recognized as a pattern. All other structural image components are summarized as details. Using other visual pattern categories, small high contrast components (<10% PW, >5% BW) are characterized as spacial details whereas low contrast components of any size (<5% BW) are characterized as intensity details. In addition, digital images contain noise of various origins. Only the high frequency noise components at the level of a few pixels are visually perceivable as random intensity fluctuations. Other low frequency noise components, which have the character of intensity details, may be recognizable only after eliminating (averaging, smoothing) the high frequency components. Several properties make noise difficult to handle and reduce. In single data sets at the level of individual pixels, noise is indistinguishable from small spacial detail lacking a dominant pattern. Noise will also disrupt the integrity of larger structural components in proportion to its strength. Thus, if visually distracting noise components are reduced by smoothing, some detail will be lost. Additionally, if the noise has a Gaussian characteristic, some noise pixels of extreme intensities will always remain as well as some low frequency components.

Image evaluation of microscopy, radiological and other similar data requires that all image details are made to be identifiable by visual pattern recognition while maintaining the spacial relation of details within the overall image. This implies that low intensity details be contrast enhanced by a factor of 10-100 and that the smallest spacial details be enlarged by a factor of 10-20. However, for correlative image component evaluation, only three image processing tasks are required.

1. Detail contrast enhancement in full frame images;

2. Digital enlargements;

3. Noise management.

In general, the evaluation of an image's information content will require a sequence of image processing steps: first, in the full frame image after contrast enhancement, details will be recognized and cross correlated with image features; then interesting, enhanced details will be enlarged and analyzed; and finally, distracting noise may be reduced. If the image contains a high noise level, the noise is first reduced to a level at which the detail information can be easily recognized. However, such image processing must fulfill some stringent conditions in order to be effective and practical. The image processing techniques must avoid processing artifacts inherent to most common image enhancement procedures which distort the spacial and most of the intensity characteristics of details. Only pixel accurate enhancement techniques promote closer visual inspection by digital enlargement. In addition, the processing must be fast (close to real time) in order not to disturb the visual recognition process; it must be able to automatically accommodate all images independent of their size, depth and content, it must be exhaustive and objective to avoid missing any existing detail; and it must be simple and without any other input than a single factor, i.e., a "visibility enhancement factor".

On the basis of its pixel accuracy, digital image processing in accordance with the present invention can provide visual access to acquired digital image data at the level of instrumental image resolution by extracting and imaging intensity defined image information classes. This data reduction fosters visual pattern recognition. However, a new strategy for visual appreciation and communication of such data must be introduced because digital image processing displays information previously not always seen and utilized and thus not a part of our visual perception and pattern experience in a particular application. The image detail information provides a visual pattern similar to the common visual repertoire and an interpretation of this pattern is surprisingly simple since it correlates with intensity contributions to individual pixels of known instrumental contrast mechanisms. Therefore, the new visual image evaluation approach directly correlates the data with all other previously gathered information available (sample characteristics, microscope or instrumental performance and contrast mechanisms) and facilitates appreciation of even the smallest information detail.

Digital image data sets are intensity maps showing the intensity at each pixel in a two dimensional array of pixels. Referring to FIG. 6, an intensity profile graphs this intensity along a line of pixels (FIG. 6, upper left box), and a series of intensity profiles outlines in a three dimensional graph the three dimensional "intensity profile" surface of the data in x and y direction. The intensities surface may vary in height and spacial extent indicating image components with certain spacial and intensity characteristics. Four different image components can be visually identified since they match basic visual perception patterns. Random intensity fluctuations over smallest distances (single pixels) are characteristic of "noise" (FIG. 6, bottom left box). In the three dimensional intensity profile surface, the noise pixels would appear as slender spikes or holes covering larger intensity components. The average intensity variation of noise (in Gaussian noise .+-.2 standard deviations {SD} similar to the typical bandwidth display of analog noise) occupies a significant intensity bandwidth (IR.sub.S) which spreads through the intensity profile surface as an evenly thick outer layer. In high-quality data, noise occupies only a few percent (<10%) of the total intensity bandwidth too small to be visually recognized. However, extracting the noise component (FIG. 6, bottom left box; upper intensity profile) and contrast stretching it to the full visual bandwidth will make the noise pixels visible and identifiable by their spacial random distribution (FIG. 6 bottom left box: bottom intensity profile and map). In some microscopic data sets, noise is always dominant, e.g., in low dose imaging or in high magnification imaging. Sometimes, lack of contrast prevents acquisition of high quality data and will result in poor data quality (noise IR.sub.s >20%).

Homogeneous intensity variations over areas of many pixels are nonrandom and present the spacial information of the data. Depending on size and intensity characteristics, three major visual pattern types can be distinguished. Small low contrast components are summarized as "spacial detail" (FIG. 6, bottom middle box). If the components are uniform they will occupy an identical significant intensity range (IR.sub.S) above the background intensity. In the three-dimensional intensity profile, the spacial details will appear as well defined small little "bumps" or "invaginations". Extraction and contrast stretching provide for easy perception of these details and recognition of the distribution pattern. In most image data sets, spacial details occupy only a fraction (IR.sub.S =2-20%) of the total intensity range. In some low magnification data, spacial details may be a dominant contrast component.

In general, small intensity variations (<5% significant intensity range) can extend over larger areas (FIG. 6, bottom right box). These components represent "intensity details" which in some data occupy only a minute fraction of the available intensity range. In the three-dimensional intensity profile, intensity details are recognizable as large flat elevations following contours of lager image components (features). Many weak contrast mechanism produce intensity details and their recognition is an important part of microscopy or other imaging services. Intensity details are the least accessible in analog imaging since the visual system cannot easily recognize large low contrast components which often do not provide shape edges. However, if extracted from the raw data and contrast enhanced, these data become strikingly accessible for image evaluation due to increased edge contrast. Linear contrast stretching will maintain proportionality between different intensity details and thus facilitate interpretation.

Finally, most image data have some large features which dominate the full frame image (FIG. 6, top right box). Their dominance comes from a nearly complete occupation of the available data intensity range. Nearly always, other details and noise are dwarfed by the feature's intensities and become hidden from visual perception since only few intensity steps remain for their accommodation. Extraction of these compressed details can recover their few intensity levels and linear contrast stretching will space the levels evenly over the full visual perception range fostering perception, pattern recognition and data appreciation. Other images may have a large intensity "background" without any intensity information which occupies a major portion of the intensity range of the data. Again, the background level reduces the image information contrast.

B. Implementation of the Image Information Enhancement Technique

The information extraction program in accordance with this invention creates an output image from two smoothed input images (or the original data set and one smoothed image). The smoothed images are preferably generated by the two-dimensional hysteresis smoothing program described above using two different cursor widths (but alternatively may be generated by a different pixel-accurate smoothing technique). The two input images must have the same number of rows and columns. The detail extraction subtracts the most smoothed image from the least smoothed image and then does linear contrast stretching on the result. In a preferred embodiment, a computer program implementing the present invention is written in assembly language for the aforementioned Visionary Systems AP.sub.x desktop massively parallel processor. It is called from a C language subroutine which runs on a PC-AT clone. The source code for both the PC-AT and the AP.sub.x is set forth in Table 3 (with the source code for the "smoothing" operation already having been set forth in Table 1). The AP.sub.x parallel processor has up to 256 16-bit processing elements (PE's) interconnected in a mesh topology. Each PE has a fast 256 word on-chip data memory and a 65536 word off-chip data memory. The PE's all have access to a shared scaler data memory and a shared program memory. A single program is executed by all the PE's in lockstep (SIMD processing). The AP.sub.x uses a PC-AT clone as a Host computer, and the AP.sub.x off-chip PE memories are memory-mapped into the Host computer's address space.

The image subtraction is done in one pass over the two smoothed images. The linear contrast stretching is performed in two passes over the output of the image subtraction. The first pass finds the minimum and maximum pixels values in the image. The second pass recalculates the output pixel values using the following equation:

where "p(x,y)" is the initial pixel value, "q(x,y)" is the final output pixel value, "minpix" is the minimum pixel value in the image, "maxpix" is the maximum pixel value in the image, and "maxval" is the maximum pixel value. In the case where "maxpix" equals "minpix", "q(x,y)" will equal zero.

An important feature of the detail enhancement technique of the present invention is the use of the aforementioned process for two-dimensional hysteresis smoothing which avoids commonly encountered processing artifacts. The "smoothing" technique is applied for noise reduction and due to its unique pixel accurate design, maintains the spacial information of the smoothed image. The image details eliminated from the smoothed image may be recovered from the original image without loss of the detail pixel accuracy. Since the contrast range of the extracted detail is limited and falls into the cursor width applied for the smoothing, the contrast enhancement of the detail information becomes predictable. This is an important advantage since it allows the extraction and enhancement for any range of image details.

A second important and novel feature of the detail extraction and enhancement technique is in its ability not only to selectively extract details but also in extracting the contrasts of larger image features by as much as its cursor width. This observation has led to the definition of the present invention as detail enhancement in full frame images since the image information eliminated through the smoothing contains both the detail information and feature information. The portion of feature information is proportional to the cursor width. On this account, the image information recovered after smoothing (through subtraction of the smoothed image from the original image) can be contrast enhanced through only one step of linear contrast stretching and maintains in principle all image information. Therefore, the detail contrasts are more enhanced than the feature contrasts which are in fact first selectively reduced. Another novel feature of the present invention is the discovery that the contrast range of all extracted information (before enhancing) is exactly known (intensity range of the image minus cursor width), thus the contrast enhancement factor is predictable and easily established.

The defined contrast range of the extracted information leads to a third feature of the detail enhancement technique of this invention, "intensity slicing". Smoothing of the original data with two different cursor sizes allows extraction of the information difference between both smoothed images. This intensity slice has the same property as the extracted details: its significant intensity range is known and it contains a proportion of the feature information dependent on the applied cursor sizes. Thus, the information contained in an intensity slice can be displayed after linear contrast stretching. As discussed with regard to FIG. 6, the development of this new process leads to a new unique definition of the information content of digital images (as containing four different intensity defined contents, i.e., noise, spacial details, intensity details, and image features/background) which will have great impact on image communication, image analysis, pattern recognition and image quantitation. The data reduction to individual intensity information classes provides a new and unique tool for the analysis of image information and quantitation of the image information content and image quality. The application of this tool in digital image processing lead to the discovery that all known contrast mechanisms of any technical imaging equipment (microscopes, telescopes, photographic and video cameras, medical imaging technologies, etc.) establish one of the three basic intensity variations found as structural classes in digital images. In principle, every image information (structural component) must be acquired and documented with contrast mechanisms which generate specific intensity differences between image components and their surrounding. The contrasts may directly image the complete structural components, part of the components (phase contrasts, diffraction contrasts) or none, requiring a complete spacial reconstruction (interference contrasts of holograms). Each of these different image contrast types can only be established in a data set as one of the three non-random information classes. Therefore, the detail enhancement technology described herein can separate and extract different contrast types for the image data and present the specific contrast information as a separate image which is easily visually recognized and objectively quantified by its significant intensity range. Since the contrast types match the visual perception and recognition parameters, such extracted contrast information is easily visualized. Connectional visual quantitation is possible only in a limited and subjective way. However, image contrast quantitation now becomes possible through the measurement of the significant intensity ranges of each contrast type. The quantitation also allows the establishment of the intensity proportion of each contrast type with the other image intensity components including noise and intensity background. This procedure provides the first objective way for image quality assessment describing the proportion and extent of the intensity components of an image. A graphic display of the intensity ranges of each of the image components in an "information cube" (FIG. 23) allows easy visual access to the image quality. Commonly, the individual "detail information contrasts" have an intensity range of only a few percent of the overall intensity range, such detail can be extracted from high precision raw data of 10-bit to 16-bit accuracy and visually displayed on an 8-bit level without compromise of the raw-data accuracy. The new information enhancement technology of this invention thus makes possible a quantitation of the image information and provides objective criteria for image quality assessment. It makes possible objective image communication via electronic networks on "high speed highways" since accurate data reduction generates clearly displayed image information without any distortions or artifacts. It allows one to visualize at the precision raw data, selected intensity information (contrast types) in deep data of 10-bit to 16-bit.

The novel detail enhancement technique of this invention utilizes two-dimensional hysteresis processing for several unique enhancement purposes which are required for the accurate enhancement of image details and which are not addressed either by the noise management application of the technique or by any other commonly used processing techniques: pixel accurate extraction of spacial details with a defined and limited intensity range, and image-accurate scaling of detail intensity and determining of a single interactively defined output parameter for the enhancement. This image accurate "enhancement parameter" is chosen independently from the image content only in relation to the desired extent of detail enhancement. The detail enhancement filter has fundamental advantages over the conventional spacial processing principles (Oho filter, Sobel filter or homomorphic filters) since it uses a different, pixel accurate intensity processing principles for the selection, extraction and enhancement and which are independent of the image content.

The following nonlimiting examples further describe the two-dimensional hysteresis smoothing and detail enhancement digital image processing techniques of the present invention.

A. System Hardware

The hardware used for image processing provides for adequate high speed processing using a 486/66 MHz PC-AT host CPU and a high level PC based AP.sub.x parallel processing technology. The array processor (AP) technology is based on single instruction/multiple data (SIMD) architecture using an expandable system of 64 to 256 16-bit processors which provide peak instruction rates of 800-3200 MIPs. The individual processors are 16-bit RISC processors which can be software configured to 32-bit mode. IEEE format single precision floating point operations are supported in 32-bit mode with peak ratings from 40-160 MFLOPs. VLSI technology allows fast one-cycle communication of 32-bit numbers. The AP.sub.x processor boards fit into PC bus slots and provide supercomputer performance. Workstations of this type are commercially available from Visionary Systems, Inc. 25 Science Park, New Haven, Conn. The standard image format is square and the image is displayed together with a simple menu on a 20 inch workstation monitor with 1280.times.1024 pixel resolution and 120 Hz refresh rate (Hitachi CM2085MU SuperScan20 monitor, allowing full stereo display with a Stereographics CrystalEyes System). The monitors are provided with custom fitted anti-magnetic Nu-metal shieldings.

Image enhancement of a typical high quality image (1K.times.1K 8-16 bit) requires 3 billion instructions per second (3000 mips). This requires supercomputer processing speed. For cost efficiency, ease of use, and portability of equipment, a PC based parallel processing system such as the aforementioned APx system constitutes an important feature of this invention. The aforementioned APx system processes 1600 mips and is therefore capable of processing an image in 2 seconds. A larger APx array will process the image much faster. A conventional smaller image (512 pixels 8 bit) is processed in less than a second. Such very short processing times (near real time) are provided only by a parallel processing system.

B. Noise Test Pattern Examples

The noise smoothing capability of the digital process imaging of this invention can be demonstrated using a test pattern derived from Gaussian noise of .+-.4 standard deviations (SD) width (1024.times.1024.times.8 bit, mean pixel value of 127, minimum and maximum pixel values of 43 and 211, standard deviation of 21, and 0.0001% clipping accuracy) (FIG. 7A). A superimposed spacial test pattern consisted of only one pixel wide features, i.e., two perpendicular double lines of 0 and 255 intensities, and two sets of small crosses of either .+-.2 SD (top half) or .+-.3 SD (bottom half) intensities (FIG. 7B: 16.times. enlarged by nearest neighbor zoom). Processing of the complete test pattern, superimposed onto the noise (FIG. 7C), with the "smoothing" technique of this invention (FIGS. 7D-F) preserved all significant spacial features and smoothed the background with increasing significance factor (cursor width) which may be set depending on the level of required processing. At a .+-.1 SD intensity level processing, all features (crosses and double lines) and noise pixels of higher/lower intensity values were maintained without distortion of their spacial integrity (FIG. 7D). All lower-frequency noise components were maintained and were recognizable in a spotty pattern. Smoothing at .+-.2 SD eliminated only the .+-.2 SD crosses and all pixels within that significant intensity range (FIG. 7E) and were dramatically improved (10 times) in their signal to noise (S/N) ratio. Finally, detail enhancement processing at a .+-.3 SD level (FIG. 7F) eliminates all noise and crosses but the double line feature. For the two latter processing examples (since the image has an even median intensity background), the noise image was first randomized prior to processing and then afterwards derandomized. Only intensity oriented "PAIP processing" in accordance with this invention can maintain the significant image content; all prior art spacial processing will corrupt the randomized data and not preserve all spacial features (FIG. 7A-H). Neither conventional "averaging" (FIG. 7G: 3.times.3 mask with center weight 1) nor conventional median filtering (FIG. 7H: 3.times.3 mask) maintained the test features' spacial integrity. Commonly, noise reduction algorithms utilize spacial masks (in the space domain or Fourier filters) and may produce spacial distortions of image details at a maximum level set by the mask or filter size. The PAIP technique of this invention maintains the spacial integrity of image details at the precision level of the raw data (pixel level) and thus is ideally suited for image processing in microscopy, radiology, astronomy or technical applications in which the primary objective is detail acquisition and evaluation.

C. Microscopy

Microscopic data evaluation concerns the recognition of contrast mechanisms which allow deduction of specific specimen properties. The raw data display such contrasts as spacial intensity variations with a characteristic significant intensity range and spacial distribution. In some cases a direct visual interpretation of contrasts is possible, in other cases reconstruction of spacial properties is required. However, in all microscopic data analyzed so far with the new intensity processing technology of this invention, the pertinent contrast information was present only in a very small significant intensity range of 1-10% of the overall intensity range of the data. The majority of raw data contained a high noise component which was as large or even larger than the high precision contrasts making intensity slicing necessary for separation of noise from pertinent spacial data. Sometimes, the relevant data was buried deep within the intensity profile. No other known image processing technology is able to recover this information automatically and produce a complete image without any spacial or proportional intensity distortions.

1. Scanning Electron Microscopy

Many microscopic data sets contain large dominant image features which occupy most of the image intensity range leaving little room for contrasts of image details at the level of the instrumental precision (single pixels). This is especially true for low magnification scanning electron microscopy (SEM). One of the superior capabilities of SEM is low magnification imaging of large specimens with high depth of field. However, at low magnification (<1,000.times.), on large objects very strong signal "collection efficiency contrasts" are generated which are more or less independent of local electron beam-specimen interactions. The presence of these detector dependent contrast mechanisms was well known and their utilization for image creation was well perfected with multi-detector systems. However, their extent could not be quantitatively assessed. Moreover, low magnification images, commonly lacking detail contrasts, were commonly used only for the communication of the large image features. Local small contrasts generated by electron beam-specimen interaction could only be utilized at higher magnifications where collection efficiency contrasts produced a signal, empty of spacial information, background reducing the image quality (signal/noise ratio). PAIP image enhancement revealed that all expected local contrasts were generated and collected at low, as well as at high, magnification but that detail contrasts were compressed in inverse proportion to the extent of the feature contrasts or background level. In addition, the visual recognition of local, small intensity variations was reduced in close proximity to large bandwidth intensity variations. PAIP image enhancement in accordance with the present invention provides an easy and fast procedure for adjustment of the proportion of detail and feature/background intensity range by interactively reducing the latter and visually evaluating the effectiveness.

(a) Conventional Scanning Electron Microscopy (SEM)

The conventional SEM image of an ant head (FIG. 8A) occupied the full intensity range (IR=0-255) and left little room for the detail contrasts produced by local electron beam specimen interaction. This high precision electron beam information was contained as spacial detail information (FIG. 8B) within a significant intensity range (IR.sub.s =1-35) of only 14% of the overall intensity range. The contrasts were comprised, distorted and reduced by signal collection efficiency contrasts of two origins. Firstly, local specimen-dependent charging phenomena modified the signal yield producing an intensity variation at the samples surface summarized as "charging contrast" (FIG. 8C) which were imaged in the intensity detail slice (IR.sub.s =35-71). Secondly, local detector-dependent "detector contrasts" (FIG. 8D) produced an overall topographic aspect of the large feature (IR.sub.s =71-256). The latter signal component was produced by backscattered electrons at the pole piece of the microscope and was well characterized for producing the overall low magnification image aspect in conventional SEM. However, the collection efficiency contrasts lead to a reduction of the detail contrast which were only recognizable on a background of middle gray levels and were fully absent in the saturated high and low intensity areas of the feature (ant head).

Image enhancement with stepwise reduction of the significant intensity range lead to a reduction of the feature contrast and to an inverse proportional increase of the detail contrasts. An over 50% reduction (FIG. 8E) maintained the major feature contrasts but visibly increased much smaller detail contrasts within the whole image preserving the overall character of the image features but also emphasized most of the image details (IR.sub.s =1-100). Notably, the signal collection efficiency contrasts were reduced and the electron probe induced contrasts were somewhat enhanced. A further reduction of the significant intensity range to 6% of the overall image intensity range (IR.sub.s =1-15) finally led to a dominance of local electron beam scattering contrasts (FIG. 8F). The image quality changed dramatically once the electron beam induced contrast mechanisms were visualized. Especially, intensity details (micro-roughness) contrasts on the specimen support and spacial details (edges along the chitinous plates covering the animal, and smallest hairs and bristles) became dominant. The enhanced image proved that the electron probe was capable of detail imaging despite an overwhelming contrast range produced by other mechanisms. Quick adjustment at the significant intensity range in the overall intensity or an intensity slice was essential for seeing what signals the electron probe was generating and thus promoted contrast interpretation.

Detail evaluation by digital magnification was facilitated if the image's intensity range was appropriately adjusted to the size of dominating contrasts. Since the raw data was nearly free of noise (IR.sub.S of less than 1%) and the spacial detail was contained in a relatively wide intensity range, detail evaluation was performed at an intensity slice of IR.sub.s =1-15. (see FIG. 8F). FIGS. 8G-H depict digitally enlarged (16.times. by nearest neighbor zoom for visualization of the individual pixels) details (just underneath the eye) of the raw data and the selected intensity range (IR.sub.S =0-15). The enhancement revealed truthfully (pixel accurate) fine structures and minute contrasts produced by the electron probe on the sample surface expanding scanning electron microscopy to scanned electron probe microscopy. Besides spacial information, small topography contrasts (relief contrasts) became identifiable. Comparison of the enhanced data with the raw data for the interpretation of structural detail was possible by flipping between both images with the stroke of one key of the keyboard. It proved that no spacial or intensity artifacts were observable even at enlargements of 10-20 times. Every detail seen in the enhanced images was present in the original data set but with lower contrasts.

(b) Medium Magnification Field Emission SEM (FSEM)

In SEM at medium magnification, a possible low magnification feature contrast converts into an empty background signal which has adverse properties similar to that of the large features, i.e., it compresses the microscopic specimen-specific contrasts inversely proportional to its size. The lack of high precision spacial information limits the optimizing of SEM application in important areas as found in industrial R&D and semiconductor failure.

For evaluation of the porosity of polypropylene filters, thin metal coating (3 nm Pt), high acceleration voltage (30 keV) and 20,000.times. magnification were required, but signal-to-noise and all image contrasts were week (FIG. 9A). An intensity slice within the intensity detail of only 2% width (IR.sub.s =21-27) reduced effectively the noise component and enhanced the material contrasts (FIG. 9B) so that the pore volume became easily quantifiable. The selective extraction of this contrast allowed the reduction of noise and charging contrasts and the full recovery of the spacial high precision information generated by the electron problem on this difficult sample.

Similar constraints are encountered at medium magnification imaging for failure analysis in the semiconductor industry. In this application, cross sections of devices were polished, slightly etched and successively coated with 3 nm of Cr. Imaging at high performance conditions (30 kV, field emission electron source) produced only weak detail contrasts (FIG. 10A: 2.times. bicubic zoom) which did not allow an identification of the device architecture. An intensity slice underneath the noise component of only 4 intensity steps (IR.sub.s =7-11) included all spacial detail of the raw data and revealed the cross section's topography in short range contrasts at the precision level of the electron beam (individual pixels) (FIG. 10B). Such enhancement is invaluable for routine FSEM application.

(c) Low Voltage Field Emission SEM (LVFSEM)

In on-line production control in semiconductor manufacturing, cross sectional line width measurements are performed routinely with low voltage field emission scanning electron microscopy (LVFSEM) since it allows imaging of non-conductors with increased over-all contrast. However, the resulting high magnification images revealed only low high resolution (short range) contrasts reduced by a background from the dramatically increased long-range signal (FIG. 11A: 2.times. bicubic zoom). In resist on polysilicon cross sections, pertinent details at the base and top of the resist lines were not evident. Despite a high noise level, fine structural detail was recovered from 5% of the over-all intensity range (IR.sub.s =13-25) excluding the noise component (FIG. 11B). Since the detail enhancement is pixel-accurate, direct manual or automatic measurement of the line width becomes possible without the need for any further image processing.

(d) Environmental SEM (SEM)

The new low vacuum SEM technology images non-conductors and wet surfaces, both often consisting of samples with low mass density and low signal yield. Higher magnification images therefore are limited by a very large noise component in addition to the common large signal background component. At medium magnification (20,000.times. and 20 kV), latex spheres on a fully hydrated plastic filter were imaged only with weak contrasts (FIG. 12A: 2.times. bicubic zoom) obscuring the interaction of the spheres with the filter's pores. Using a narrow intensity slice (IR.sub.s =13-23) and additional digital zooming (lower right image corner) allowed an evaluation of the filter properties (FIG. 12B: 4.times. bicubic zoom). In order to preserve the ultra-structural details of the filter pores (double zones), some of the high intensity noise component had to be included in the extracted contrast information.

(e) High Resolution Field Emission SEM (HRFSEM)

The new ultra-high resolution FSEM instruments provide a 0.5 nm probe diameter and are capable of working at magnifications of 100,000-1,000,000.times. in order to take advantage of the high theoretical resolution. In practice, contrast quality sets the limitation at high magnification. At high magnification, even in these in-lens microscopes, a large signal background is generated which compresses the high precision short-range contrasts (spacial detail information). High magnification cryo-imaging (100,000.times., 30 kV, sample temperature -120.degree. C.) of a molecular preparation on thin C film (2 nm thickness), shadowed with a 1 nm continuous Cr film, produced only disappointingly low contrasts with no molecular details being visible (FIG. 13A). After digital zoom and intensity slicing beneath the noise component, all expected high precision contrasts were revealed (IR.sub.s =5-15), i.e., all small structures (2-5 nm in size) were imaged with a distinct shadow produced by the high resolution contrasts of the aimed-for SE-I signal component (FIG. 13B). The small size of the SE-I contrasts was unexpected but indicative for the large background produced here even under optimal imaging conditions (very thin specimen, low-Z metal coating, cryo-imaging for contamination reduction).

2. Transmission Electron Microscopy Images

TEM contrasts are a good example for the superposition of various contrast mechanisms and the difficulty in visual contrast information interpretation in the space domain (the image). In the TEM at the level of smallest structural details, both phase contrast and scattering contrast occur and are superimposed. Phase contrasts can be easily analyzed and reconstructed in the Fourier domain using the transfer function of the optical system and microscope imaging parameters. Low-dose imaging is a prerequisite for beam sensitive materials but increased noise and lack of contrast hinders immediate evaluation of acquired raw data. In addition, the image quality may be obscured by beam damage, insufficient dose, and other factors (contamination, instabilities etc.).

(a) Cryo-Transmission Electron Microscopy (Cryo-TEM)

Cryo-TEM is of utmost importance in biomedical electron microscopy. It is the high-resolution microscopy of choice for the imaging of the ultrastructure of cells and tissues. In a further advance, energy filtering is applied for a reduction of the signal background which obscures and compresses the contrasts of interest. Since in cryo-TEM on vitrified sectioned materials, no staining is applied, only weak phase contrasts are available for high resolution information. However, the contrasts do not present a coherent spacial information but image parts of the spacial frequencies in positive and other parts in negative contrast according to the transfer function of the particular imaging conditions (lens aberrations, illumination and refocusing). The phase contrast transfer function relates a specific intensity characteristic to each of the spacial frequencies. This inherent intensity characteristic of the TEM contrasts can be determined by the processing technology described in this application and used for a separation and imaging of each of the existing particular spacial frequencies of the raw data. Even if the energy filtering is applied for generating an enriched phase contrast data set, other signal component with the same energy range are included and produce a large background signal (up to 80% of the total signal as can be shown here). Intensity background reduction through intensity slicing maintains the spacial character of the data and generates enhanced phase contrast images enriched in certain spacial frequencies. This imaging technology is especially important for high resolution cryo-TEM in phase contrast on frozen-hydrated biological materials. The original low-dose energy-filtered TEM raw data set of a frozen-hydrated unstained cryo-section of an apple leaf (25,000.times. instrumental magnification) revealed uneven contrast distribution exhibited by a large background signal (FIG. 14A). Most of the overall intensity range (IR=0-232) was occupied by scattering contrasts of large image features. Small structural details had low phase contrast and were barely visible and buried in noise. Lack of contrast hindered quality assessment. Conventional contrast enhancement by histogram equalization did not improve the phase contrast image quality because it also enhanced all other intensity variations in a non-spacial manner and may lead to a disintegration of detail components or loss of intensity details, i.e., loss of some intensity values which were summed into bins of especially inappropriate intensity levels.

The different spacial phase contrast component were enriched in different successive intensity slices proving the coherence of the intensity profile and validating the new image information processing technology. Interactive visual control facilitated the determination of a significant intensity level at which noise reduction and preservation of fine structural detail were optimized. Intensity slicing (IR.sub.s =25-35) just underneath the noise component revealed the smallest spacial phase contrast components (FIG. 14B) and successive slicing (IR.sub.s =35-45) extracted larger spacial phase contrast components (FIG. 14C). Finally, the largest spacial frequencies of this data set were found buried well in the depth of the intensity profile (IR.sub.s =55-75). The latter component contained the knife marks (originating from the cryo-sectioning) as well as ice crystal contamination. Separation of the phase contrasts or its various components will facilitate the reconstruction of the specimen's ultrastructure.

(b) High Voltage TEM (HVTEM)

High voltage microscopy provides for high resolution but is limited by a reduction of image contrast. Therefore, often dark field microscopy is used for a recovery of some of the low contrasts components. The extent of collected fine structural information in such high precision data was unknown but can be assessed through intensity slicing. A 300 kV TEM image (60,000.times. magnification in conical dark field illumination) of mineral platelets coated with surfactant reveals only few fine structural details (FIG. 15A: 2.times. zoom) seen in some darker areas in between plates (holes in the surfactant bilayer, 5-10 nm thick, formed in between the plates). A narrow intensity slice (IR.sub.s =3-15) revealed underneath the noise components all contrasts produced by the surfactant bilayers in between the plates as well as on top of the plates (surfactant monolayers). The majority of the intensity data were occupied by scattering contrasts from the mineral plates and a signal background. The very high precision obtainable with high voltage TEM was, although not visible, present in the raw data. This finding will dramatically improve the application of these expensive research tools.

(c) High Resolution TEM (HRTEM)

Highest, atomic resolution can be achieved in high voltage TEM at very high acceleration voltages on very thin specimens. However, phase contrasts and scattering contrasts always superimpose and obscure the desired high resolution information. In an evaporated gold island preparation at 5,000,000.times. magnification, atoms of the gold as well as the supports were imaged but obscured by mass inhomogeneities of the sample (FIG. 16A). Intensity slicing (IR.sub.s =3-35) of the phase contrasts revealed a coherent image of the atomic ultrastructure of the sample (FIG. 16B). Digital zoom (Insert: 4.times.) allowed easy first evaluation of the image quality. Reconstruction of these data will be required for a spacial data output of highest precision.

3. Scanning Transmission Electron Microscopy Images

Scanning transmission electron microscopes are important instruments in R&D due to their ultra high resolution capability, high depth of field and limited demand for specimen preparation. However, the imaging capabilities of these instruments are limited as well as other microscopies by the proportion of high precision, short range contrasts and their intensity background.

(a) Scanning Transmission Electron Microscopy (STEM)

The analysis of STEM data (2,300,000.times. magnification) of small (only a few nm large) grains of high-Z metal in a low-Z matrix (FIG. 17A: 2.times. zoom) was limited by the uneven sample thickness (producing the background signal). High precision contrast information was found in a narrow intensity slice (IR.sub.s =5-35). The enhanced data revealed all particles present and facilitated particle quantitation. The intensity background was produced by the extensive electron scattering in the thicker portions of the sample. However, conventional background intensity compensation would not be effective in preserving the particles differential contrasts. Only intensity slicing as described in this application can reveal the particles differential contrast as evenly high (and undistorted) in the entire image.

4. Scanning Probe Microscopy Images

The new scanning probe microscopes promised acquisition of high precision data from all specimens but in practical applications their use is limited by a lack of contrast at the expected precision level. In tact, the microscope's precision is the highest available for any surface microscope, measuring in fractions of one Angstrom, but such short-range contrast information is not visible in images of rough surface found in routine applications.

(a) Atomic Force Microscopy (AFM)

A typical data set is presented by polymer coating films (800,000 magnification) formed from 20 nm small flat particles on a Teflon surface (FIG. 18A: 2.times. zoom). The few Angstrom high substructures of the film surface were barely recognizable in the raw data. Since the AFM contrasts are proportional to the height of image components, the particulate substructure contrasts covering the whole film surface was compressed to a few intensity levels. Intensity slicing below the noise (IR.sub.s =7-13) revealed within only 17 intensity levels the small high precision contrasts and generated a detailed image adequately representing the expected AFM image quality.

5. Light Microscope Images

Light microscope contrasts were always accompanied by large background components which derived from light scattering in inhomogeneous samples. A method for reduction of this contrast limiting background was introduced through the confocal LM. However, also CFLM data find a general limitation in noise as well as the small intensity range of short range contrasts.

(a) Light Microscopy (LM)

The bright field image of paper fibers, suspended in water, at medium magnification (400.times.) was dominated by adsorption contrasts while the short range contrasts, resulting from optical properties, were barely visible (FIG. 19A). Contrast enhancement of this high precision information was easily established by a simple intensity slice beneath the noise component at 10% of the full intensity range (IRs=7-23). Through this processing, all expected refraction, diffraction and polarization contrasts became visible (FIG. 19B); although it was thought that this precision information is available only in certain circumstances.

(b) Confocal Laser Light Microscopy (CFLM)

The confocal laser light imaging technology aims for a reduction of the signal background produced from light scattering with the sample. However, low signal yields and high noise levels are characteristic for the CFLM data. As in other microscopy data, noise management is important and can be facilitated through the intensity slicing since random and non-random (structural) information can easily be visually recognized and assessed. An image of an in-situ injection of a fluorescent lipid dye into a living photo receptor cell revealed little detail (FIG. 20A). The extent of the image noise component was fast determined by several intensity slices, i.e., FIG. 20B: upper noise component (IR.sub.s =1-17), FIG. 20C: median noise component (IR.sub.s =17-37). All structural data (FIG. 20D) were found below the large noise component (IR.sub.s =1-57:22% IR) within only 15 intensity levels (IR.sub.s =57-72). The data slice was easily found by slicing through the intensity profile and evaluating in each slice the proportion of random and non-random information. The ability to "see" the noise component, and to evaluate its penetration into the fine structural data is of utmost importance for microscopy since it allows one to visually assess the practical intensity threshold which separates noise and structural information at minimized loss of detail information.

D. Medical Imaging

Of special interest is digital image enhancement for medical image data evaluation. Although Radiologists are well trained for visual evaluation of image documented on photographic films, detail and pattern recognition and evaluation is restricted by the individual visual abilities and the excess of data beyond the visual limitations. New digital medical image sensors acquire images at 12-bit or 16-bit level and a resolution better than films. New digital mammograms provide 10K.times.12K.times.12 bit-data (PrimeX, Carlsbad, Calif.). Visual evaluation of such large data at the precision level of acquisition is not possible and will require digital image processing technologies. The described intensity slicing technology provides the only strategy for the evaluation of the data information without any compromise in precision. The following two examples were digitized from film.

1. Mammography

Mammogram evaluation assesses the tissue structure of the mammary gland. Two stages of tumor growth pattern can be distinguished: 1. Early indications are seen in widening of gland ducts and in micro-calcifications composed of groups of small high contrast deposits, 2. Late indications are seen in growth of tissue masses in round areas of increased contrast (more water content from dense cell accumulations). The structural diagnostic criteria fall into the two basic image information classes of image details, i.e., spacial details (smallest contrast variations as found in fibers and micro-calcifications) and intensity details (large area contrasts as found in tissue components and alterations). Problems in mammogram evaluation arise from excessive tissue density and concomitant superposition of contrasts. A dense mammogram (FIG. 21A) was evaluated by digital PAIP image information enhancement. Very little detail is visible in the original raw data image. As in the previous data sets, all detail information was imaged within a small significant intensity range of only 12%. In a routine screening of the X-ray film at a light table, the analog image evaluation must be completed within less than 1.5 minutes. The composite detail image (FIG. 21B: IR.sub.s =1-21) allowed a correlation of both detail information, the spacial detail (FIG. 21C: 1-11) and the intensity detail (FIG. 21D: IR.sub.s =11-21). The spacial detail image allowed an easy recognition of the tissue fibers and of a group of individual micro-calcifications (FIG. 21C: circle). Other high intensity detail may arise from superimposition of fibrous components which could be easily identified. The intensity detail image revealed the gland ducts and large round areas of decreased contrast indicative of fat depositions (less water content). Nearly all gland ducts can be followed leading to a point of origin (nipple). Some ducts were found dramatically increased in width (FIG. 21D: circle).

However, in traditional mammograms, detail information may be inaccessible due to structural super-position. This phenomena can be reduced through stereoscopic three-dimensional (3-D) imaging. New high sensitivity digital mammogram sensors will make it possible to take two tissue exposures under altered angle of view (stereo angle of 4.degree.-6.degree. without much increased risk of excessive X-ray exposure. The PAIP filter of this invention is fully equipped for stereo imaging of detail enhanced and enlarged image details. It is ideally suited for stereo imaging since it produces images of tissue details with high contrast and high edge sharpness, a prerequisite for visual 3-D perception.

The PAIP evaluation of spacial and intensity details takes less than one minute in automatic grid search made at four times enlargement. It is much less straining than analog image evaluation (aided with a magnifying glass) since the image information is presented in an optimal visual fashion. The easy recognizable display of all detail information insures that data interpretation is no longer limited by lack of visual access to the data but only by limitations in data quality or diagnostic experience of the investigator which in addition could now be supplemented through image communication via electronic networks.

2. Chest X-Ray

Many medical data sets contain large dominant image features which occupy most of the image intensity bandwidth leaving little room for contrasts of image details. This is especially true for chest and other X-rays which image bones and tissue together. Local contrasts are constrained within only a few percent of the image intensity range. PAIP image enhancement in accordance with the present invention revealed that all expected local contrasts were generated and collected but that detail information contrasts were compressed in inverse proportion to the extent of the background contrasts. In addition, the visual recognition of local, small intensity variations was reduced in close proximity to large intensity variations.

PAIP image enhancement provided an easy and fast procedure for adjustment of the proportion of detail and feature intensity ranges by interactive reduction of the latter while visually evaluating the processing effectiveness. A chest X-ray image (FIG. 22A) occupied the full intensity range and left little room for the detail contrasts which had a significant intensity range of only 5%. They were only recognizable in part within the soft tissue background and were fully absent in the high and low intensity areas of the X-ray. Detail enhancement with stepwise reduction of the significant intensity range lead to a reduction of the feature intensity range and to an inverse proportional increase of the detail contrasts. An over 90% reduction (FIG. 22B: IR.sub.s =1-10) still maintained the major feature contrasts but visibly increased much smaller detail contrasts within the whole image. A further reduction of the significant intensity range to 5% of the overall image intensity range finally led to a dominance of local contrasts (FIG. 22C: IR.sub.s =1-5). The image quality changed dramatically once the sensitive small X-ray adsorption contrasts were visualized. In high intensity areas at 1% significant intensity level, mostly noise was imaged (FIG. 22D IR.sub.s =1-3), bottom right) but finest spacial details were preserved in dark areas. Such a typical non-linear noise component can be easily reduced though processing using a non-linear enhancement.

E. Quantitation of Digital Image Information

The definition in digital images of any origin and content through PAIP of defined contrast classes and the quantitation of these classes provides a unique and new tool for image quality quantitation. The four intensity information classes can be schematically represented in an information cube (FIG. 23) which depicts the relative significant intensity (z coordinate) over the image (x and y coordinates). Each information class is presented by a simplified three-dimensional intensity profile and stacked upon each other in the sequence of access through PAIP processing; at the top is the image noise, followed by the spacial detail, the intensity detail, and at the base is the image feature/background. The latter is shaded if its spacial information is limited (light gray) or absent (background: dark gray). In FIG. 23 info-cubes are presented of some of the microscopy data sets depicted in FIGS. 8-20. As additional information, the information classes are labelled with symbols along the left vertical axis, the total intensity range (IR) of the data set it indicated in the lower left corner, and the file name and the percentage range of each of the information classes is given in sequence to their appearance in the data. It is evident, that each image shows an individual proportion of its information components, but that common to all images is a reduction of visually perceivable "image quality" when the proportion of the feature component is more than 50% of the maximum relative significant intensity range. In FIG. 23, the data are arranged with decreasing image quality. Another application of the image quality assessment is found in image evaluation at the time of acquisition. This is important in microscopy, where constantly varying contrast conditions as well as degradation of the specimen during microscopy hinder visual data assessment. In FIG. 23 at the very bottom two image data characteristics are shown for a low quality data acquisition and subsequent improvement of image quality through adjustment of microscopic parameters. Although the size of the noise component remained nearly unchanged, the proportion of the spacial and intensity detail was dramatically changed by a factor of 4-10 affecting both image information components differently. These examples indicate the power of the new digital image processing technology of this invention which may be applied to all possible digital image applications due to its intensity processing character, accuracy, objectivity and automation capability. ##SPC1##

While preferred embodiments have been shown and described, various modifications and substitutions may be made thereto without departing from the spirit and scope of the invention. Accordingly, it is to be understood that the present invention has been described by way of illustrations and not limitation.