Method and apparatus for zooming digital images

[0002] Not applicable.

[0003] Not applicable.

[0004] The present invention relates to imaging methods and apparatus and more specifically to a method and apparatus which facilitates magnification or zooming of a portion of an image on a display.

[0005] Digital imaging is extremely important in many different applications. For example, digital imaging has proven invaluable in the medical imaging art where vast amounts of data are collected and used to generate images for observation on a video monitor or display. Although not so limited, in the interest of simplifying this explanation the present invention will be described in the context of medical imaging. A typical display includes a two-dimensional raster of pixels. For the purpose of this explanation, although raster pixels may be arranged in any of several different configurations, it will be assumed that pixels are arranged in distinct rows and columns.

[0006] To generate images on the display a processor collects all image data and generates intensity signals or pixel values for each display pixel. The pixel values are provided to a video driver which excites each pixel separately in accordance with an associated pixel value. From a distance the contrast between display pixel intensities is seen as an image.

[0007] Magnification of digital images in real time is needed in many different applications. For example, it may be advantageous to magnify a section of a medical image which includes a tumor. To this end, software has been developed which allows a user to select an image section for magnification and then magnifies the selected section.

[0008] One problem with image magnification has been selecting pixel intensities in a magnified image which reflect the initial image and result in a clean magnified image. For example, assume an intensity range between 0 and 100 where first and second adjacent pixel values correspond to intensities 55 and 92. Also assume that upon magnification, the area corresponding to the first and second pixels increases and covers 32 pixels (e.g. 16 pixels each). In this case, if the 16 magnified pixels corresponding to the first pixel and the 16 magnified pixels corresponding to the second pixel are provided with intensities 55 and 92, respectively, and all other magnified pixels are similarly magnified, the result is a highly granulated image which, in many cases, is not useful for the intended purpose of closer examination.

[0009] Instead of exciting pixels in the magnified image as indicated above (i.e. with either intensities 55 or 92), other solutions have been adopted by the industry with varying success and at varying costs. The most common methods for calculating new pixel intensities are nearest neighbor, bilinear and bicubic interpolation methods.

[0010] According to the nearest neighbor method, when points corresponding to first and second adjacent pixels on an initial image are separated by magnification so that the initial image points correspond to third and fourth pixels which are separated by a plurality of other pixels, the intensity of each of the other pixels (e.g. pixels between the third and fourth) are set equal to the intensity of the closest of the third or fourth pixel. This solution has the advantage of being computationally simple and therefore can be implemented easily using existing imaging hardware. Unfortunately, nearest neighbor methods only increase magnified image quality slightly and therefore are unacceptable for many applications.

[0011] Bilinear methods generally linearly fill in pixel intensities. For instance, in the example above where points corresponding to first and second adjacent pixels on an initial image are separated by magnification so that the initial image points correspond to third and fourth pixels which are separated by three other pixels and the first and second pixel intensities were 55 and 92, respectively, the other pixel intensities are linearly determined and are approximately 64, 73 and 83. High speed bilinear image magnification is now commonly available in accelerated graphics hardware. Unfortunately, while this solution generates a better image than the nearest neighbor methods, this solution requires much more processor time to perform necessary computations and still does not provide an extremely accurate magnification.

[0012] Bicubic interpolation methods generally take into account the intensities of more than just first and second adjacent pixels when determining the intensities of pixels in a magnified image which are between image points which correspond to the initial first and second pixels. In effect, these interpolation methods mathematically identify pixel intensities on one or more curves wherein the curves correspond to proximate initial pixel intensities of several pixels about an area.

[0013] These interpolation methods are extremely accurate and generate diagnostic quality magnified images. Unfortunately, these methods require massive amounts of processor time and therefore, in many cases, cannot be performed in real time because of processor limitation. For this reason high quality bicubic image magnification is not generally available.

[0014] One way to speed up calculations is to provide special hardware which is specifically designed to perform specific calculations. For example, many image processing systems include special hardware to perform either one or two dimensional high speed convolution filtering and linear or bilinear interpolation processes required in many imaging application. Unfortunately, hardware solutions have not yet been provided to facilitate bicubic interpolation.

[0015] Therefore, it would be advantageous to have a method and an apparatus which can be used with existing hardware to facilitate image magnification wherein resulting magnified images are of a quality which is essentially identical to the quality achievable using bicubic interpolation and wherein the method facilitates real time magnification.

[0016] It has been recognized that specific hardware which already exists in many imaging work stations can be used to imitate bicubic interpolation thereby providing extremely accurate image magnification. Importantly, while the functions and calculations performed by the hardware are extremely computationally intensive, the special hardware can perform the functions and calculations extremely quickly thereby giving the appearance of real time zooming or magnification.

[0017] Generally, the hardware includes a convolver, an interpolator, and a processor for identifying convolver coefficients. After the coefficients are identified, the convolver convolves pixel intensities for each pixel in the original image region to be magnified thereby generating at least two intermediate values. The interpolator interpolates the intermediate values to generate output pixel intensities.

[0018] Specifically, according to a one dimensional method of zooming, the invention includes a method and apparatus to be used with a digital imaging system to facilitate magnification of a region of interest on an initial digital image wherein the region of interest includes a plurality of pixels of interest, adjacent halves of adjacent pixels of interest forming interpixel intervals of interest. Each pixel of interest is characterized by a pixel intensity. The system includes convolution filter and interpolation hardware. The method comprises the steps of, for each interpixel interval of interest, identifying adjacent pixels, for each adjacent pixel, identifying a convolution window and for each window, determining a convolution filter coefficient for each pixel in the window, convoluting the pixel intensities of the pixels in each window using the coefficients to generate an intermediate value, and interpolating the intermediate values to generate interpolated pixel intensities, the interpolated pixel intensities together forming an interpolated image.

[0019] Preferably, according to the one-dimensional method, the initial pixel array is arranged in y columns and x rows and the step of selecting includes the step of, for each interpixel interval of interest between pixels in adjacent columns, selecting first and second row convolution windows consisting of pixels which are in the same row as the interval of interest and wherein, the method further includes the steps of, for each interpixel interval between pixels in adjacent rows on the interpolated image, selecting first and second separate column convolution windows on the interpolated image, each column window including adjacent image pixels within the same column as, and proximate the interpixel interval of interest, for each window, determining a convolution filter coefficient for each pixel in the window and convoluting the pixel intensities in each window using the coefficients to generate an intermediate value, and interpolating the intermediate values to generate a final pixel intensity, the final pixel intensities together forming the final image. The invention also includes a two-dimensional method and apparatus for zooming which is similar to the one-dimensional method and apparatus.

[0020] In the description, reference is made to the accompanying drawings which form a part hereof, and in which there is shown a preferred embodiment of the invention. Such embodiment does not necessarily represent the full scope of the invention and reference is made therefore, to the claims herein for interpreting the scope of the invention.

[0021] FIG. 1 is a schematic view illustrating one dimensional convolution interpolation methods according to the present invention;

[0022] FIG. 2 is a is a schematic illustrating a convolution window in accordance with a second embodiment of the present invention;

[0023] FIG. 3 is a schematic view illustrating a relationship between intermediate intensity values and all pixel intensities according to the second embodiment to the invention;

[0024] FIG. 4 is a perspective view of a computer system used to facilitate the inventive methods;

[0025] FIG. 5 is a schematic diagram of hardware used in both the first and second embodiments to the present invention;

[0026] FIG. 6 is a schematic illustrating hardware according to a one dimensional convolution embodiment of the present invention;

[0027] FIG. 7 is a schematic diagram illustrating a region of interest and expanded region of interest according to the present invention;

[0028] FIG. 8 is similar to FIG. 6, albeit illustrating hardware required for a two-dimensional convolution system according to the present invention;

[0029] FIG. 9 is a flow chart illustrating operation of the hardware of FIG. 5;

[0030] FIG. 10 is a flow chart illustrating operation of the hardware of FIG. 6; and

[0031] FIG. 11 is a flow chart illustrating operation of the hardware of FIG. 8.

[0032] A. Theory

[0033] Generally, it has been recognized that bicubic interpolation can be closely estimated or mimicked by performing a two-dimensional convolution of initial pixel intensities to generate intermediate values and then linearly interpolate between intermediate values to generate specific final image pixel intensities. While such convolution and interpolation procedures are computationally intensive, many imaging systems now include hardware which can perform required computations extremely quickly such that computations appear to be performed in real time. Specifically, the hardware includes convolution filters and linear or bilinear interpolators which can quickly provide the functions required to estimate pixel intensities which would be generated via software based bicubic interpolation.

[0034] Some systems only include hardware which can be used to perform either one or two-dimensional convolution and either linear or bilinear interpolation. The present invention includes a first method referred to herein as a one-dimensional convolution (ODC) method which can be used by hardware which is limited to one-dimensional convolution and linear interpolation to estimate output pixel intensities. In addition, the present invention also includes a second method referred to herein as a two-dimensional convolution (TDC) method which can be used by hardware which can perform two-dimensional convolution and bilinear interpolation to estimate output pixel intensities. Each of the one and two-dimensional convolution methods is described separately below.

[0035] In each of the methods below it will be assumed that display pixels are arranged in x horizontal rows and y vertical columns. In addition, it will be assumed that a command received from a user indicates that a specific image region should be magnified by a factor n where n is 5 (i.e. magnification is by a factor of 5). Thus, generally, the area of a first image which corresponds to a single pixel will be magnified such that the same area in the final image corresponds t o a b lock of pixels having 5×5 dimensions. The region to be magnified will be referred to herein as a region of interest and each pixel within the region of interest will be referred to as a pixel of interest.

[0036] 1. One-Dimensional Convolution (ODC) Method

[0037] According to the ODC method, when a command is received to magnify region of interest by a factor of n, prior to convolution and interpolation an expanded region of interest which includes the region of interest and an additional number of rows above and below the region of interest is selected. The additional rows in the expanded must be processed in the x dimension so that data is generated for processing in the y dimension. This requirement will become more clear below.

[0038] After the expanded region has been selected, convolution and interpolation hardware is used to modify pixel data first in the x (i.e. row) dimension to generate an interpolated image. Thereafter, the interpolated image data is fed again to the convolution and interpolation hardware which modifies the pixel data in the y (i.e. column) dimension thereby generating a final or output image. Each pass through the hardware includes a two step process corresponding to each pixel processed during the pass. In the first step, for each pixel in the image being processed the convolution hardware selects a separate convolution window on the image being processed (i.e. on the initial image during the first pass and on the interpolated image during the second pass). During the first pass each window (i.e. a row window) is in the same row as a corresponding pixel of interest and during each second pass each window (i.e. a column window) is in the same column as a corresponding final pixel. In addition, each window includes pixels adjacent the pixel of interest. The number of pixels in each convolution window depends on a filter size m where m is typically a number of pixels (e.g. 5, 7, etc.). Hereinafter m will be assumed to be 5 unless indicated otherwise.

[0039] To better understand convolution windows and the inventive ODC method refer to FIG. 1 wherein six initial pixels P1 through P6 are illustrated. The pixels P1 through P6 are to be magnified by a factor of 5. For the purpose of this explanation, when referring to the ODC method, the term “interpixel interval” will be used to refer to adjacent halves of adjacent pixels. For example, in FIG. 1, an interpixel interval corresponding to pixels P3 and P4 is identified in phantom as P34. Similarly, there is an interpixel interval between each other two adjacent pixels. To magnify pixels P1 through P6, according to the ODC method each interpixel interval is magnified by a factor of 5. To this end, two convolution windows including m pixels are selected which correspond to each interval. For example, where a filter size m is 5, for interval P34, the windows are first window 10aincluding pixels P1 through P5 and second window 10bincluding pixels P2 through P6. The intensities of pixels P1 through P6 are identified as I₁through I₆.

[0040] When the 5 times magnification request is received, in effect, the area in the initial image corresponding to each interpixel interval is first expanded in the x dimension to cover five pixels. For example, in FIG. 1, the area corresponding to interpixel interval P34 is expanded into interpolated pixels P1′ through P5′ having intensities O1 through O5 forming an output pixel block 12. Similarly, the interval areas between other initial adjacent pixel pairs P1 and P2, P2 and P3, P4 and P5 and P5 and P6 are each expanded into five distinct interpolated pixel intensities. Thus, pixels P1 through P6 are first expanded in the x dimension to form 25 separate interpolated pixels (i.e. 5 output pixels corresponding to each interpixel interval), each output pixel characterized by a separate intensity. Thereafter, the same procedure is repeated in the y dimension to provide five separate output pixel intensities for the area between any two adjacent interpolated pixels.

[0041] Referring still to FIG. 1, to expand interval P34 in the x dimension, convolution windows 10aand 10bare first selected about adjacent pixels P3 and P4. After convolution windows 10aand 10bhave been selected the convolution hardware convolutes each of the windows 10aand 10bseparately, thereby generating two intermediate values b₁and b₂, respectively. The convolution equations for intermediate values b₁and b₂are expressed as: $\begin{array}{cc}{b}_1=\sum _{i=1}^m{a}_i{I}_i\mathrm{and}& \left(1\right)\\ b_2=\sum _{i=1}^m{a}_i{I}_{i+1}& \left(2\right)\end{array}$

[0042] where a_iare coefficients of the convolution filter which depend on the magnification factor n, where, in this example, n is 5.

[0044] After each of intermediate values b₁and b₂have been identified, the intermediate values are provided to the linear interpolation hardware which interpolates intensities b₁and b₂across n output pixels. To this end, the interpolation hardware performs the following equations to determine intensities O_kof final pixels P1′, P2′, P3′, P4′ and P5′ where k is 1 through n, in the present example 1 through 5. Where n is odd: $\begin{array}{cc}{O}_k=\frac{\left(n-k+1\right)}{n}{b}_1+\frac{\left(k-1\right)}{n}{b}_2& \left(3\right)\end{array}$

[0045] and, where n is even: $\begin{array}{cc}{O}_k=\frac{\left(n-k+0.5\right)}{n}{b}_1+\frac{\left(k-0.5\right)}{n}{b}_2& \left(4\right)\end{array}$

[0046] Substituting values b₁and b₂from Equations 1 and 2 into Equations 3 and 4 and simplifying yield, where n is odd: $\begin{array}{cc}{O}_k=\frac{\left(n-k+1\right)}{n}\sum _{i=1}^m{a}_i{I}_i+\frac{\left(k-1\right)}{n}\sum _{i=1}^m{a}_i{I}_{i+1}& \left(5\right)\end{array}$

[0047] and where n is even: $\begin{array}{cc}{O}_k=\frac{\left(n-k+0.5\right)}{n}\sum _{i=1}^m{a}_i{I}_i+\frac{\left(k-0.5\right)}{n}\sum _{i=1}^m{a}_i{I}_{i+1}& \left(6\right)\end{array}$

[0048] Equations 5 and 6 can be expressed as: $\begin{array}{cc}{O}_k=\sum _{i_1=1}^{m+1}\sum _{i_2=2}^m{\lambda }_{i_1{i}_2}{a}_{i_2}{I}_{i_1}& \left(7\right)\end{array}$

[0049] where the λ_iivalues are a function of the coefficients with corresponding I values from either Equation 3 or Equation 4, depending on if n is odd or even. For example, if n is odd and i₂is 1, λ_iiis from Equation 3 or: $\begin{array}{cc}{\lambda }_{i_{1}1}=\frac{n-k+1}{n}& \left(8\right)\end{array}$

[0050] The only unknowns in Equation 7 are convolution filter coefficients a_i. Therefore, if coefficients a_i2can be identified for a specific magnification factor n, Equations 1, 2 and 3 in the case of an odd factor n or, in the case of an even factor n, Equations 1, 2 and 4, can be performed to determine intensities O1-O5 of pixels P1′-P5′.

[0051] To identify the convolution filter coefficients a_i, a set of equations similar to Equation 7 is first derived in accordance with conventional bicubic interpolation methods. Thereafter common coefficients of Equation 7 and the cubic Equations are set equal and solved for the convolution filter coefficients. To this end, as well known, where there are four equidistant points x₁, x₂, x₃and x₄on an image and a function f defines intensity values I₁, I₂, I₃and I₄of the four points, respectively, the intensity value of the function at some point x between points x₂and x₃can be determined as follows. First, assume dx is the normalized distance defined by: $\begin{array}{cc}\mathrm{dx}=\frac{x-x_2}{x_3-x_2}& \left(9\right)\end{array}$

[0052] with distance dx so defined, the following cubic coefficients c₁through c₄can be calculated: $\begin{array}{cc}{c}_1=-\frac{1}{3}\mathrm{dx}+\frac{1}{2}{\left(\mathrm{dx}\right)}^2-\frac{1}{6}{\left(\mathrm{dx}\right)}^3& \left(10\right)\\ c_2=1-\frac{1}{2}\mathrm{dx}-{\left(\mathrm{dx}\right)}^2+\frac{1}{2}{\left(\mathrm{dx}\right)}^3& \left(11\right)\\ c_3=\mathrm{dx}+\frac{1}{2}{\left(\mathrm{dx}\right)}^2-\frac{1}{2}{\left(\mathrm{dx}\right)}^3& \left(12\right)\\ c_4=-\frac{1}{6}\mathrm{dx}+\frac{1}{6}{\left(\mathrm{dx}\right)}^3& \left(13\right)\end{array}$

[0053] The intensity value O_kof function f at point x, using cubic interpolation can therefore be expressed as: $\begin{array}{cc}{O}_k=f\left(x\right)=\sum _{i=1}^4{c}_1{I}_1& \left(14\right)\end{array}$

[0054] Comparing Equations 7 and 14, and equating coefficients with corresponding input pixel intensities I_i, a set of up to four linear equations for the convolution filter coefficients a_ican be expressed as: $\begin{array}{cc}\sum _{i_2=1}^m{\lambda }_{{\mathrm{ii}}_2}{a}_{i_2}=c_i& \left(15\right)\end{array}$

[0055] All four cubic coefficients c; can be determined by solving Equations 10 through 13 and, therefore, the only unknown values in Equation 15 are the five coefficients a₁through a₅.

[0056] To generate additional equations for determining coefficients a₁through a₅, the procedure above is repeated for each of the output pixels. Thus, referring again to FIG. 1, because n is 5 and there are therefore five output pixels P1′ through P5′ in block 12, four c_ivalues are generated for each separate pixel P1′ through P5′ to generate twenty c_ivalues and appropriate equations like Equation 15, a separate equation for each of the twenty c_ivalues. Thus, twenty (i.e. 4n) cubic equations are derived with only 5 (i.e. m) unknowns (i.e. a₁through a₅).

[0057] There can be less than four equations if the pixels obtained by bilinear interpolation used by the convolution filter do not have any contribution from some input pixels I_i. In this case, the number of equations is reduced by the number of such pixels.

[0058] The system of 4n equations can be expressed in matrix form as:

[0059] C=ΛA  (16)

[0060] where C is the vector of corresponding coefficients c_ifrom right sides of equations 15, Λ is the matrix of coefficients λ_ijof those linear equations, and A is a vector of m filter coefficients a_i. For practical values of zoom factors n and filter sizes m the matrix of Equation 16 represents an over determined system of linear equations. In this case, the solution to matrix Equation 16 is determined by using the linear least squares method which takes the form:

[0061] A=(Λ^TΛ)⁻¹Λ^TC  (17)

[0062] The linear least squares method and in particular Equation 17 are well known in mathematics and therefore will not be explained here in detail. For a general explanation of the linear least squares method reference should be had to any standard algebra text which teaches matrix algebra.

[0063] Upon determining the coefficients aj corresponding to the convolution filter that is to be applied, the whole procedure for zooming digital images is defined.

[0064] 2. Two-Dimensional Convolution (TDC) Method

[0065] According to the TDC method, when a command is received to magnify a region of interest by factor n, a two dimensional hardware convolution filter and a bilinear interpolator are used to convolve and interpolate initial pixel intensities simultaneously in both the x and y dimensions thereby generating a final output image.

[0066] Thus, unlike the ODC method which first processes values in one dimension and then processes values in a second dimension to magnify, the TDC method processes in both dimensions simultaneously. Other than the TDC method processes in both dimensions simultaneously. Other than this distinction and more complex mathematics which are required for simultaneous two-dimensional processing, the TDC method is very similar to the ODC method.

[0067] For the purpose of this explanation, when referring to the TDC method the term “interpixel interval” will be used to refer to adjacent quarters of four adjacent pixels. Referring to FIG. 2, the general pixel scheme for applying the TDC method to a region of interest is illustrated. The initial region of interest includes a pixel block of (m+1) (m+1) pixels. In FIG. 2, I₁₁through I_(m+1)(m+1)denote source image pixel intensities corresponding to pixels of interest P₁₁through P_(m+1)(m+1), bag through b₂₂denote intermediate values obtained by applying a 2D convolution filter to intensities I₁₁through I_(m+1)(m+1), numeral 20 outlines a convolution filter window including pixels P₁₁through P_mmand O₁₁through O_nnrepresent intensities of output pixels P₁₁′ through P_nn′ obtained by bilinear interpolation between intermediate pixels b₁₁, b₁₂, b₂₁, and b₂₂. Generally intensities I₁₁through I_mmwill be referred to as I_ij.

[0068] Intermediate values b₁₁through b₂₂are obtained by applying the m×m convolution filter to the input pixels I_ijaccording to the following expressions: $\begin{array}{cc}{b}_{11}=\sum _{i=1}^m\sum _{j=1}^m{a}_{\mathrm{ij}}{I}_{\mathrm{ij}}& \left(18\right)\\ b_{12}=\sum _{i=1}^m\sum _{j=1}^m{a}_{\mathrm{ij}}{I}_{\mathrm{ij}+1}& \left(19\right)\\ b_{21}=\sum _{i=1}^m\sum _{j=1}^m{a}_{\mathrm{ij}}{I}_{i+1j}& \left(20\right)\\ b_{22}=\sum _{i=1}^m\sum _{j=1}^m{a}_{\mathrm{ij}}{I}_{i+\mathrm{ij}+1}& \left(21\right)\end{array}$

[0069] where a_ijare the coefficients of the two-dimensional convolution filter.

[0070] After intermediate values b₁₁through b₂₂are determined bilinear interpolation is applied to generate output intensities O₁₁through O_nn. FIG. 3 illustrates the relationship between intermediate values b₁₁, b₁₂, b₂₁, and b₂₂and the output pixels O₁₁through O_nn. Each of the output pixels O_ijoccupies a different location relative to the intermediate values b₁₁through b₂₂. All the other output pixels generated by the linear interpolation occupy the same relative locations to some intermediate pixels. The values of the output pixels O₁₁through O_nn, with respect to the intermediate values are given by the following equations: $\begin{array}{cc}\begin{array}{c}{O}_{\mathrm{ij}}=\frac{\left(n-i+1\right)\left(n-j+1\right)}{n^2}{b}_{11}+\frac{\left(n-i+1\right)\left(j-1\right)}{n^2}{b}_{12}+\\ \frac{\left(i-1\right)\left(n-j+1\right)}{n^2}{b}_{21}+\frac{\left(i-1\right)\left(j-1\right)}{n^2}{b}_{22}\end{array}& \left(22\right)\end{array}$

[0071] for i, j=1, . . . , n (n odd), and $\begin{array}{cc}\begin{array}{c}{O}_{\mathrm{ij}}=\frac{\left(n-i+0.5\right)\left(n-j+0.5\right)}{n^2}{b}_{11}+\\ \frac{\left(n-i+0.5\right)\left(j-0.5\right)}{n^2}{b}_{12}+\\ \frac{\left(i-0.5\right)\left(n-j+0.5\right)}{n^2}{b}_{21}+\frac{\left(i-0.5\right)\left(j-0.5\right)}{n^2}{b}_{22}\end{array}& \left(23\right)\end{array}$

[0072] for i, j=1, . . . , n (n even).

[0073] When the expressions for intermediate values b₁₁through b₂₂from Equations 18 through 21 are substituted in Equation 22, the following expression is obtained: $\begin{array}{cc}\begin{array}{c}{O}_{\mathrm{ij}}=\frac{\left(n-i+1\right)\left(n-j+1\right)}{n^2}\sum _{k=1}^m\sum _{l=1}^m{a}_{\mathrm{kl}}{I}_{\mathrm{kl}}+\\ \frac{\left(n-i+1\right)\left(j-1\right)}{n^2}\sum _{k=1}^m\sum _{l=1}^m{a}_{\mathrm{kl}}{I}_{\mathrm{kl}+1}+\\ \frac{\left(i-1\right)\left(n-j+1\right)}{n^2}\sum _{k=1}^m\sum _{l=1}^m{a}_{\mathrm{kl}}{I}_{\mathrm{kl}+1l}+\\ \frac{\left(i-1\right)\left(j-1\right)}{n^2}\sum _{k=1}^m\sum _{l=1}^m{a}_{\mathrm{kl}}{I}_{k+1l+1}\end{array}& \left(24\right)\end{array}$

[0074] which can be expresses as: $\begin{array}{cc}{O}_{\mathrm{kl}}=\sum _{i_1=1}^{m+1}\sum _{i_2=1}^{m+1}\sum _{i_3=1}^m\sum _{i_4=1}^m{\lambda }_{i_1{i}_2{i}_3{i}_4}{a}_{i_3{i}_4}{I}_{i_1{i}_2}& \left(25\right)\end{array}$

[0075] where λ_iiiiare the constant coefficients from Equations 22. Similar expressions are obtainable using Equations 23.

[0076] As well known in the art, using known expressions for bicubic interpolation alternate expressions for pixel values O_klcan be obtained which can be collectively expressed as: $\begin{array}{cc}{O}_{\mathrm{kl}}=\sum _{i=1}^4\sum _{j=1}^4{c}_{\mathrm{ij}}{I}_{\mathrm{ij}}& \left(26\right)\end{array}$

[0077] where c_ijare the coefficients derived from the expressions for bicubic interpolation in the case of magnification by n. Comparing Equations 25 and 26 and equating coefficients with corresponding input pixel values I_ij, a set of up to sixteen linear equations for the filter coefficients a_iiare obtained which can be collectively expressed as: $\begin{array}{cc}\sum _{i_3=1}^m\sum _{i_4=1}^m{\lambda }_{{\mathrm{iji}}_3{i}_4}{a}_{i_3{i}_4}=c_{\mathrm{ij}}& \left(27\right)\end{array}$

[0078] There can be less than sixteen equations if the pixels obtained by bilinear interpolation used by the convolution filter do not have any contribution from some input pixels I_ij. In that case, the number of equations is reduced by the number of such pixels. Repeating this procedure for all of the output pixels O_ij(i, j=1, . . . , n) a system of (16n) linear equations for the convolution filter coefficients is obtained which can be expressed in matrix form as:

[0079] C=ΛA  (28)

[0080] where C is a vector of the corresponding coefficients c_ijfrom right side of Equations 26, Λ is the matrix of coefficients λ_ijklof those linear equations, and A is the vector of m×m filter coefficients a₁₁through a_mm. For practical values of zoom factors n and filter sizes m by m this represents an over determined system of linear equations. In this case, the coefficients a_ijcan be determined using the linear least squares method indicated by Equation 17 above:

[0081] Upon determining the coefficients a_ijof the convolution filter that is to be applied, the whole procedure for zooming digital images is defined.

[0082] B. Hardware and Operation

[0083] Referring now to FIG. 4, a conventional workstation 30 is illustrated which includes a computer 32, a display 34 linked to the computer 32, and two interface devices, a keyboard 36 and a mouse 38. Keyboard 36 and mouse 38 allow a user to provide information to the computer including commands to control the computer 32. Information including images, is displayed on display 34 for observation. For the purpose of this explanation it is assumed that an image is displayed via display 34 and that a workstation user uses either board 36 or mouse 38 to select a region of interest on the initially displayed image to be magnified by a factor of 5 (i.e. n=5).

[0084] In addition to other processing hardware, computer 32 must include hardware capable of either one-dimensional convolution filtering and linear interpolation or two-dimensional convolution filtering and bilinear interpolation. To this end, referring to FIG. 5, computer 32 includes an interpolation coefficient calculator 40, a cubic coefficient calculator 42 and a linear least square determiner 44. Referring also to FIG. 9, at block 160, calculator 40 receives magnification factor n and uses factor n to determine interpolation coefficients. To this end, on one hand, where a hardware based convolver is only capable of one dimensional convolution and a hardware based interpolator is only capable of linear interpolation, calculator 40 uses the coefficient expressions from Equations 5 and 6 to determine interpolation coefficients at block 162. For example, where n is 5 and m is 6 the four equations corresponding to Equation 15 above for the first output pixel (i.e. n=1) are:

[0085] λ₁₁a₁+λ₁₂a₂+λ₁₃a₃+λ₁₄a₄+λ₁₅a₅=c₁  (29)

[0086] λ₂₁a₁+λ₂₂a₂+λ₂₃a₃+λ₂₄a₄+λ₂₅a₅=c₂  (30)

[0087] λ₃₁a₁+λ₃₂a_2+λ33a_3+λ34a₄+λ₃₅a₅=c₃  (31)

[0088] λ₄₁a₁+λ₄₂a₂+λ₄₃a₃+λ₄₄a₄+λ₄₅a₅=c₁  (32)

[0089] Looking at Equation 5 for I, where i is 1 corresponding to Equation 29, there is only a single coefficient expression for a₁, the expression being (n−k+1)/n and therefore coefficient λ₁₁is known to be (n−k+1)/n and all other coefficients λ₁₂, λ₁₃, λ₁₄and λ₁₅are known to be zeros. Where i is 2 corresponding to Equation 30, referring to Equation 5, coefficient λ₂₁is (k−1)/n, coefficient λ₂₂is (n−k+1)/n and all other λ coefficients in Equation 30 are zero. Where i is 3 corresponding to Equation 31, coefficient λ₃₁is zero, coefficient λ₃₂is (k−1)/n, coefficient λ₃₃is (n−k+1)/n and all other coefficients λ in Equation 31 are zeros. Where i is 4 corresponding to Equation 32 coefficients λ₄₁and λ₄₂are zeros, coefficient λ₄₃is (k−1)/n, coefficient λ₄₄is (n−k+1)/n and all other coefficients are zeros. The k value used in each of the coefficients corresponds to an output pixel and Equations like 29 through 32 are examined for each intended output pixel thereby generating 4n interpolation coefficients λ which are provided to determiner 44.

[0090] On the other hand, where the hardware convolver is a two dimensional convolver and a bi-linear interpolator is provided, at block 162 calculator 40 uses Equations 22 or 23, depending on if factor n is odd or even, to determine interpolation coefficients λ_ijklwhich are provided to determiner 44.

[0091] Referring still to FIGS. 5 and 9, at block 164 cubic coefficient calculator 42 also receives factor n and determines bicubic coefficients. Where the system convolver facilitates only one dimensional convolution calculator 42 uses Equations 9 and 10 through 13 or other similar equations to determine coefficients c₁through c₄for each separate output pixel. Where the system can perform two dimensional convolution, calculator 42 generates sixteen coefficients c₁through c₁₆for each output pixel using conventional bicubic equations which are well known in the art. Thus, 16n coefficients are generated and provided to determiner 44.

[0092] In addition to receiving interpolation coefficients λ and cubic coefficients c, determiner 44 also receives a value m indicating the filter dimensions. At blocks 166 and 168 determiner 44 forms the matrix equation indicated by Equation 17 and solves the equation to generate m or m×m coefficients a which are required in either Equations 1 and 2 or in Equations 18, 19, 20 and 21. Coefficients a_ior a_ijare provided as an output from determiner 44.

[0093] The remaining hardware in the case of the ODC method is illustrated in FIG. 6 while the remaining hardware in the case of the TDC method is illustrated in FIG. 7. Referring to FIG. 6, to facilitate the ODC method computer 32 (see FIG. 5) further includes a 1D convolver 56, a linear interpolator 58 and a data buffer 60.

[0094] Referring also to FIG. 10 at block 180, convolver 56 receives each of the initial image pixel intensity I_iwhich corresponds to the expanded region of interest. Referring also to FIG. 7, a region of interest 80 (i.e. region of an image to be magnified) includes A rows and B columns. To convolve and interpolate convolution windows for each interval in the region of interest 80 in the x dimension (i.e. horizontal), the windows include m+1 pixel intensities corresponding to m+1 pixels which are adjacent, and in the same row as, the interval of interest. Thus, for intervals on the lateral edges of region 80 corresponding windows will extend laterally from the region of interest by as much as m/2 pixels. One such laterally extending window is identified by number 82. Pixels which are laterally positioned with respect to the region of interest 80 are similar to intervals within region 80 in that they have not yet been convoluted and interpolated. Therefore, convolution windows which laterally extend during the first pass of an image through hardware include only non-convoluted pixels.

[0095] However, after intervals within a region of interest have been expanded in the x dimension, pixels in rows directly above and below the region of interest have not been expanded. Thus, upon the second pass of the image through the hardware to expand intervals in the y dimension, for intervals near the top and bottom of the region which has already been expanded in the x-dimension, suitable convolution windows cannot be selected. For example, for an interval in the top row of an x-dimension expanded region, when a convolution window including 5 pixels is chosen, while the pixels below the interval of interest and within the window have been expanded in the x-dimension like the interval of interest, the pixels above the interval of interest have not been similarly expanded in the x-dimension and therefore cannot be suitably used to convolute and interpolate the interval of interest.

[0096] To overcome this problem, referring to FIG. 7, the region of interest 80 is expanded to include m/2 rows of pixels above and m/2 rows of pixels below the initial region of interest resulting in expanded region 84. Thereafter, after expansion in the x dimension, expansion in the y dimension can be performed on intervals in rows corresponding to the original region of interest using convolution windows for top and bottom located pixels which extend into region 84 above region 80 and similarly into the region below region 80.

[0097] Referring still to FIGS. 6 and 10, in addition to receiving intensity signals I_ifor each pixel within the expanded region 84, at block 180 convolver 56 also receives convolution filter coefficients a_ifrom determiner 44 (see FIG. 5), filter size value m and factor n. Convolver 56 first selects pixel pairs adjacent each interval of interest and then selects a separate convolution window for each adjacent pixel on the expanded region of interest, and, at block 182, forms Equations 1 and 2 generating intermediate values b₁and b₂for the windows.

[0098] At block 184 interpolator 58 receives values b₁and b₂and factor n and solves either Equation 3 or Equation 4 (depending on if factor n is odd or even) for each of k output pixels to determine pixel intensities O₁through O_k. Intensities O₁through O_kare determined for each interval in the original region of interest 84 (see FIG. 7) and together intensities O₁-O_kdefine an interpolated image (i.e. the interpolated region expanded in the x dimension). The interpolated image is temporarily stored in buffer 60. Once the interpolated image has been completed, that image, including all pixel intensities, is provided to convolver 56 a second time. At block 186, for every interval between adjacent pixels in adjacent rows of the interpolated image which corresponded to rows in the initial region of interest 80 (see FIG. 7), convolver 56 identifies the adjacent pixels and selects a convolution column window for each adjacent pixel and convolves the intensities in the windows according to Equations 1 and 2 generating intermediate values b₁and b₂. At block 188, interpolator 58 receives values b₁and b₂, performs either Equation 3 or Equation 4 (depending on if n is odd or even) and generates final output intensities which together define the final and magnified image. Computer 32 (see FIG. 5) uses intensities O_kto drive display 34 to display the magnified region of interest.

[0099] Referring now to FIG. 9, the remaining hardware to facilitate the TDC method includes a 2D convolver 90 and a bi-linear interpolator 92. Referring also to FIG. 11, at block 150 convolver 90 receives the image pixel intensities I_icomprising the region of interest, the convolution filter coefficients a_i, factor n and filter value m. Convolver 90 selects four adjacent pixels for each interpixel interval and then selects a separate convolution window for each interpixel interval within the region of interest. At block 152 convolver 90 performs Equations 18 through 21 to determine intermediate values b₁₁, b₁₂, b₂₁and b₂₂. At block 152 interpolator 92 receives each of values b₁₁through b₂₂and factor n for each interval in the region of interest and interpolates each set of four values b₁₁through b₂₂to generate nxn pixel intensities O_ijby solving either Equation 22 or Equation 23, depending on if factor n is odd or even. Pixel intensities O_ijdefine the final output magnified image and can be used by computer 32 (see FIG. 5) to display the magnified image on display 34.

[0100] It should be understood that the methods and apparatuses described above are only exemplary and do not limit the scope of the invention, and that various modifications could be made by those skilled in the art that would fall under the scope of the invention.

[0101] To apprise the public of the scope of this invention, we make the following claims.