Method and apparatus for indexing and retrieving images from an images database based on a color query

[0002] This application is related to U.S. applications respectively entitled METHOD AND APPARATUS FOR REPRESENTING COLORED SURFACES VIA A COLOR CODE BOOK, U.S. Ser. No. 09/440,711, filed Nov. 16, 1999, and METHOD AND APPARATUS FOR DETECTING REGIONS BELONGING TO A SPECIFIED COLOR SURFACE IN AN UNSEGMENTED IMAGE, now U.S. Pat. No. 6,469,706 B1, both hereby incorporated by reference.

[0003] 1. Field of Invention

[0004] The present invention relates generally to the field of image storage databases. More specifically, the present invention is related to recognizing and retrieving target images from a database of images.

[0005] 2. Discussion of Prior Art

[0006] The representation and discussion of color and color perception usually involves first defining a color space or color model. A color space (or model) comprises a plurality of attributes which are used to specify the variety of colors we perceive. Different colors are distinguished by having different attribute values.

[0007] One color model, or color space, is not absolutely better than another. They each have been designed from different perspectives and thus have different qualities, as well as handicaps, which make them each suitable for different applications.

[0008] RGB color space, probably the best known color model, is related to the hardware capabilities of a CRT which uses an electron gun to excite red, green, and blue phosphors to produce perceivable color. One difficulty with the RGB model is that it is quite sensitive to the variation in color values brought about by lighting changes. Intensity is distributed throughout all three parameters of this model, thus-creating color values which are highly sensitive to scene brightness. In essence, the effect of a change in the illumination intensity of a scene is approximated by a multiplication of each RGB color by a uniform scalar factor.

[0009] A simple example of RGB color space's unstableness with respect to illumination is the gross change in perceived color which accompanies changing the brightness value on a computer monitor. Other color spaces have been developed which are more stable with respect to intensity variations. The HSV model is one such example.

[0010] Physiologically, human eyes tend to respond to and evaluate colors according to hue (H), saturation (S), and luminance (L). Hue refers to color produced by different wavelengths of light. Saturation refers to the measure of a pure hue in a color. As an example, red and pink are of similar hue but have different saturation levels. Brightness, luminance, or intensity is the measure of the amount of light reflected or emitted from an object. Because the luminance value is separate from H and V, this model shows more stability than the RGB model with respect to illumination differences. However, representation of color in the HSV model is not totally immune to illumination changes.

[0011] A number of color spaces have been developed, and will continue to be developed, to address the various applications which rely on color information to accomplish their goals. One useful undertaking, which is well known in the art, is the formulation of transform operations to map one color space into another. Using these transforms, a scene or image created for one application is appropriately mapped to a second color space for an application which benefits from the image being in the second representation.

[0012] One imaging application that is receiving a lot of attention is that of a large collection of images stored in a searchable database. In particular, querying such a collection to locate images which contain desired objects requires detection, recognition and indexing of color surfaces or objects within an image.

[0013] In the past, image content detection and recognition have relied on controlled imaging conditions of the objects of interest. Background surface color, illumination source, illumination quality, illumination amount, object orientation, shadowing and occlusion were all carefully controlled to ensure that proper recognition of an object was possible.

[0014] However, the ability to detect and recognize color surfaces or objects in images, across a wide range of imaging conditions, is fundamental to the creation of robust and useful content-based image retrieval systems. Image databases typically include images from a plurality of sources under a variety of imaging conditions. Consistent representation of a particular color surface throughout these varying conditions is necessary to enable color queries (i.e. queries for a particular color surface). The present invention introduces a color code book to provide an imaging-condition invariant representation of color surfaces present within an image. The prior art discussed below fails to provide for such a robust, compact and simple method of representing color surfaces.

[0015] As explained previously, color represented through three filtered channels (e.g. RGB) of a typical camera-grabbed input is notoriously unstable with respect to changes in imaging conditions. Prior art attempts at providing robust and accurate color representations include mapping images into a perceptually uniform color space (e.g. Lab, Munsell, etc.) and representing color as a function of surface reflectance. While attaining some limited success in improving color representation, both these methods require pre-segmentation of an image before color surface recognition analysis can occur in that image. Some prior art methods do avoid a pre-segmentation step but, as a result, require some simplifying approximations to be made such as assuming all surfaces are flat (e.g. Retinex theory) or that surfaces are linear combinations of only a few basic reflectances (e.g. Maloney and Wandel).

[0016] Image segmentation is generally recognized as the partition of an image into a set of non-overlapped regions, homogenous with respect to some criteria, whose union covers the entire image.

[0017] In the prior art described below, the examples of region localization and recognition methods all require segmentation of an image prior to subsequent analysis. Automated segmentation of images is very computationally and time intensive and often introduces errors due to its imprecision. Manual image segmentation, which is frequently more accurate, is typically labor intensive and cost prohibitive. Segmentation is a difficult and time consuming step because of the variety of artifacts that exist in images. Correctly determining where an object is and what an object is are both made difficult by object occlusion, illumination artifacts, shading, reflections, object orientation, and blending with a similar background or texture.

[0018] The present invention, unlike the prior art discussed below, enables the localization and recognition of image regions, or color surfaces, without first requiring pre-segmentation of an image.

[0019] The patent to Takahashi et al., (4,203,671) provides for a method of detecting, within RGB space, whether or not an image region is human skin under various conditions.

[0020] The patent to Gast et al. (4,414,635) provides for a method of labeling regions according to the color of pixels within the region, but utilizes a predetermined light source for imaging.

[0021] The patent to Hoffrichter, et al. (4,623,973) provides for a color recognition method which transforms an RGB pixel value of a pre-isolated color region into a luminance-chrominance space.

[0022] The patent to Huntsman (4,884,130) provides for a new color space which describes colors in a way more closely related to human color perception.

[0023] The patent to Lugos (4,917,500) provides for an object recognition system which only works on pre-segmented images under predetermined lighting conditions.

[0024] The patents to Inuzuka et al. (5,307,088) and Morag et al. (5,517,334) provide for a method of quantitizing and transforming pixel colors from one color space to another. Recognizing and locating color surfaces do not appear to be discussed.

[0025] The patent to Davis et al. (RE 33,244) provides for a color transparency printing system which dynamically adjusts color look-up tables in response to non-linearities in the components of the system. However, no optimization of the actual color space to assist with surface recognition is provided.

[0026] The patent to Washio et al. (5,109,274) provides for an image processing system which determines the chromatic and achromatic nature of an image in order to implement appropriate algorithms for that image type.

[0027] The patent to Fossum (5,220,646) provides for a computer graphics system which implements a z-buffer such that colors of rendered images and background colors can be determined and hidden lines removed in a single memory pass. While applicable to polygon images, this system does not appear to consider color surface detection or coding.

[0028] The patent to Shibazaki (5,386,483) provides for a method of iteratively examining image regions to determine pixel colors to assist with choosing overlapping regions in order to achieve high printing quality.

[0029] The patent to Kasson et al. (5,390,035) provides for mapping a 3-dimensional input color domain into an m-dimensional color space. The conversion method is directed towards generating a function which improves accuracy while minimizing computational burden; color surface differentiation does not appear to be discussed.

[0030] The patent to Shamir (5,568,555) provides for an encoding system which increases symbol density by utilizing both color and color intensity as variables. Linear analysis is performed to calibrate color intensities from different colorant sources; however, a color code book training system or analysis of free-form images is not provided.

[0031] The patent to Ring et al. (5,754,184) provides for an image fidelity system which converts images from a device-dependent space to a device-independent space.

[0032] The IBM TDB entitled “Color Graphics Picture Segmentation” describes a method of compressing images which segments neighborhoods of pixels that have similar colors that are separated by pixels having a different color. Each pixel is mapped to one of N allowable colors thus segmenting the image into N color regions.

[0033] The IBM TDB entitled “Presentation Space Coloring” describes an object-oriented method of generically specifying color which is independent of the underlying system data types and resolution. While not color-space dependent, this scheme does not consider combining color spaces or optimizing a color space to aid with surface recognition.

[0034] Whatever the precise merits, features and advantages of the above cited references, none of them achieve or fulfill, individually or in combination, the purposes of the present invention. Specifically, they fail to provide for a compact and simple coding of color surfaces which provides robust representation of the surfaces under varying imaging conditions. Without such a color surface code book, the prior art also fails to provide for an image region localization and recognition method that can be performed on an unsegmented image. In addition, the prior art fails to show an image database indexing method which relies on a color surface code book to provide the ability to perform color queries.

[0035] Indexing and retrieving images from an image database, in response to content-based color queries, is accomplished utilizing semantic labels and codes from a color code book. The color code book entries provide descriptions of a variety of colored surfaces under varying imaging conditions. Images from a database are first analyzed to determine which images contain which color surfaces. For each color label in a color code book, an index entry is created linking the label with the image regions located during image analysis. The index allows queries based on semantic labels to identify similar images in the database. In addition, querying by sample images are also possible by first determining which color surfaces (and their corresponding labels) are present in the sample image.

[0036] FIG. 1 illustrates a method to form a color code book according to the present invention.

[0037] FIG. 2 illustrates a method to detect and label color regions within an image according to the present invention.

[0038] FIG. 3 illustrates a method of indexing an image database, according to the present invention, in order to retrieve images by semantic color queries.

[0039] FIG. 4 illustrates a method of retrieving images, according to the present invention, in response to queries by images.

[0040] While this invention is illustrated and described in a preferred embodiment, the device may be produced in many different configurations, forms and materials. There is depicted in the drawings, and will herein be described in detail, a preferred embodiment of the invention, with the understanding that the present disclosure is to be considered as a exemplification of the principles of the invention and the associated functional specifications of the materials for its construction and is not intended to limit the invention to the embodiment illustrated. Those skilled in the art will envision many other possible variations within the scope of the present invention.

[0041] In order for a color description scheme to be useful in such diverse areas as content-based image retrieval, color fidelity during scanning, printing, transmission and rendering of color images, image recognition, and surface localization, the scheme must be robust to changes in imaging conditions such as illumination changes, illumination artifacts, occlusions and pose changes.

[0042] As described earlier, the typical RGB color space of digital cameras and frame-grabbers is particularly unstable in its description of color with respect to changes in imaging conditions. While earlier methods to improve color stability have been proposed that involve mapping RGB color space into a more stable representation,the present invention uses an optimal color space and creates a color code book, which provides a simple and compact representation of color surfaces present in a collection of images.

[0043] The current invention provides a method of representing colored surfaces in a manner which is resistant to changes in illumination of the surface and changes in the surface's appearance (e.g. pose, orientation). A multistep method is illustrated in FIG. 1 which generates a color code book that captures, in a simple and compact form, the description of color surfaces under different imaging conditions.

[0044] The first step 102 is to obtain training samples of color surface classes under a number of lighting conditions and of various surface class members. The training samples are referred to as patches and are a variety of representative images of a particular color surface class.

[0045] For a particular color surface class (e.g. skin, sky, grass, etc.), a number of images of sample patches are acquired under well chosen illumination conditions. Examples of some, but not all, lighting conditions are 1) natural outdoor lighting including morning, mid-day and evening, 2) artificial outdoor lights (e.g. floodlights),3) natural indoor lighting, including sunlight filtering through windows, curtains, and skylights, and 4) artificial indoor lights including tungsten, fluorescent, household light bulbs, etc. The more patches that are acquired, the more accurate the color surface recognition and localization will be, due to the more robust surface description that results.

[0046] In addition, the sample surface patches acquired include different members of that particular surface class. As an example, if the surface class is skin, then images of Caucasian, Black and Asian skin, as well as other races, are acquired also.

[0047] The underlying goal when acquiring sample patches is to select them in such a manner so as to increase statistical variation within and among the patches. In a preferred embodiment, therefore, patches are acquired of non-flat portions of color surfaces, patches are approximately 40×40 pixels, and for colors where there are many shades (e.g. oranges and reds) more sample patches are acquired than for colors having fewer shades (e.g. blues). These specific criteria, while not absolutely required, work together to increase the patches' statistical variations. Typically, 9 to 10 sample patches per surface are sufficient to provide the necessary variation and no requirement exists that each class have the same number of patches nor that each sample patch have the same size.

[0048] Conventional digital cameras and frame-grabber hardware is preferably used to acquire the sample patches. Typically, these devices represent the image in RGB color space. Other methods of acquiring images and other color spaces, however, are also contemplated to be within the scope of the present invention.

[0049] For each sample patch image acquired in the previous step, the pixels are mapped 106 to a redundant color space of at least four dimensions formed by combining multiple conventional color spaces 104. Furthermore,the pixel is mapped in such a way as to ignore its luminance component.

[0050] A number of different ways of mixing conventional color spaces to form the target redundant color space are possible. As an example, YES space (ignoring the Y component), a common Xerox™ encoding standard, combined with Ohta and Kanade space (also ignoring the Y component) provides a 4-dimensional color space which ignores luminance.

[0051] While this example is a preferred redundant space, other color space combinations and even those of higher dimensions than 4 are contemplated within the scope of the present invention. As an example, YES and YIQ is another combination which provides a 4 dimensional redundant space once the Y components are ignored. In general, linear combinations of RGB derived spaces, again ignoring any luminance component, are combinable to form higher-dimensional, redundant color spaces. Non-linear spaces such as CIE-XYZ, CIE-xy Y, CIE-uv Y, etc. require knowledge of the illuminant being used and do not combine well to form a redundant space.

[0052] The next series of steps 108 perform Fisher discriminant analysis (FDA) to separate and characterize the colored surfaces of interest. The goal of FDA is to derive a transformation of vector space which decreases the overlap of class distributions and thus increases the correct classification rate. As applied here, the goal is to create a color space such that when pixels from an unknown surface are mapped into it, the likelihood of correctly determining which color surface class the pixels are from is maximized. The general concepts of cluster analysis and discrimination are known; however, the present invention's application of these principles to color spaces is novel. A full description of those elements as applied specifically to color surfaces is provided. A rigorous proof of each mathematical step, however, is not necessary for a complete understanding of the present invention.

[0053] To aid in referring to the different objects being manipulated in this analysis each color surface is designated as a class.

[0054] Each pixel, x[t], in the redundant color space is considered to be a multidimensional vector, particularly 4D in the preferred embodiment. A matrix, Sc[c], is defined as the unnormalized class covariance matrix and is given by the following sum over all x[t] which belong to a class (i.e. color surface) c.

[0055] Sc[c]=Σ(x[t]−u[c])(x[t]−u[c])^T

[0056] Where u[c] is the mean vector of the c-th class.

[0057] A within-class scatter matrix, WSc, is obtained by combining the class-specific scatter matrices Sc[c] for all classes:

[0058] WSc=ΣSc[c]

[0059] A between-class scatter matrix, BSc, is obtained by the following summation over every class c:

[0060] Bsc=Σn[c](u[c]−u)(u[c]−u)^T

[0061] Where u is the total mean vector of all x[t] and n[c] is the number of vectors x[t] for the c-th class.

[0062] There exists a matrix W which maximizes the quotient:

[0063] J(W)=|W^TBScW|/|W^TWScW|

[0064] The columns of the optimal matrix W are the generalized eigen vectors of the following expression:

[0065] BScw[i]=L[i]WScw[i]

[0066] Where L[i] is the i-th eigen value and w[i] is the i-th column of W and corresponds to the i-th eigen vector. In a preferred embodiment, the first (largest) eigen vector, V, is determined and removed from the equation which is then re-optimized to find the second largest eigen vector V′. Alternatively, choosing V′ to be the eigen vector corresponding to the second largest eigen value during the original optimization often provides sufficient discrimination between surface classes. While using other eigen vectors is an alternative, the larger the eigen value, the more discriminative the direction represented by the eigen vector. The remaining eigen vectors contribute little to discrimination and their inclusion can actually hurt performance since they essentially represent noisy components of the original signal.

[0067] The two eigen vectors chosen, typically denoted V and V′, are combined to form an optimal color space 108 into which the pixels, x[t], are transformed.

[0068] The next step 110 of creating the surface color code book, involves orthogonally projecting all pixels, from the redundant color space, into the optimal color space. One result of the projection is that the surface representation is pose invariant since no image position information is retained in the color description.

[0069] Within the optimal color space, clusters of pixels are formed by transforming pixels from the sample patches into this new space. There is not a one-to-one correspondence between sample patches and clusters; similar sample patches may map into the same cluster within the optimal color space. As an example, within the class “skin”, sample of Asian skin under similar lighting conditions will likely map into the same cluster. The notation used is: there are j color surface classes (e.g. skin, sky, grass, etc.) and for each class there are i clusters.

[0070] A mathematical representation for each cluster of pixels is now calculated to form the color code portion of the color code book. For each cluster of pixels within the optimal color space, a mean, M_ij, a covariance matrix, A_ij, eigen values, γ_ij, and eigen vectors, Ψ_ij, are calculated. In particular, it is noted that the different clusters from a particular color surface are of elliptic shape and have a commonality of direction (or orientation) that remains largely unchanged under changes in appearance due to illumination and other imaging conditions. However, the location of the cluster in a color space does change with illumination and thus choosing a range of illuminations results in modeling a surface as a collection of parallel clusters in color space. The clusters, being elliptic-shaped, are completely represented by their mean and covariance matrix (including eigen values and eigen vectors). One novel aspect of the present invention is that the optimal color space, formed from {V, V′}, varies according to the color surface classes that need to be distinguished. Particular to the present invention, this means that a color code book, optimized for a particular set of likely-to-be-encountered color surfaces, is used to improve surface classification decisions during subsequent image analysis operations.

[0071] Next, each color surface class is assigned 116 a semantic label, C_j(e.g. skin, grass, sky, etc.) and associated 114 with all its different cluster representations to create a set:

[0072] <C_j,{(M_1j, A_1j, γ_1j, Ψ_1j), (M_2j, A_2j, γ_2j, Ψ_2j), . . . }>

[0073] Finally, the color code book is formed by combining 118 all the different class representations: $\begin{array}{c}\{⟨V,V^{\prime }⟩,\{⟨C_1,\{\left(M_{11},A_{11},{\gamma }_{11},{\Psi }_{11}\right),\left(M_{21},A_{21},{\gamma }_{21},{\Psi }_{21}\right),\dots \}⟩,\\ ⟨C_2,\{\left(M_{12},A_{12},{\gamma }_{12},{\Psi }_{12}\right),\left(M_{22},A_{22},{\gamma }_{22},{\Psi }_{22}\right),\dots \}⟩,\dots \}\}\end{array}$

[0074] Note that the color code book has a color space, {V, V′} associated with it implicitly.

[0075] Many different specific manners of internal storage of the color code book in a computer system are contemplated within the scope of the present invention. A preferred storage arrangement uses the color code book as a type of index table with entries consisting of semantic labels and short multi-digit codes. These codes, or labels, are then used as an index to the actual, multi-parameter representations of the color surfaces.

[0076] Thus a surface color code book is formed which 1) describes a color surface using a compact code which is invariant to illumination and pose changes, 2) is formed in an optimal color space which varies based on all the color surfaces being distinguished, and 3) represents each surface color class by a list of clusters described by their mean, covariance, eigen values and eigen vectors.

[0077] As previously discussed, one benefit of a color code book as described above is that detection and recognition of colored surfaces within images is accomplished without the need to pre-segment the image.

[0078] FIG. 2 illustrates the steps of a method used to locate regions in an image corresponding to a particular surface and assigning an appropriate label, from the color code book, to those regions.

[0079] The first step, after acquiring a subject image 202, is to project every pixel in an image to be analyzed into the color space of the color code book 204. As with the training patches described earlier, the pixel must first be mapped into a redundant color space. In many image database applications, a commonality exists between the surfaces likely to be present in the given images. In the preferred embodiment, therefore, a code book and its accompanying color space, is selected according to the color surfaces anticipated. However, the present invention considers, within its scope, other criteria for selecting color code books.

[0080] The closest color is then assigned to each pixel. To determine the closest color, the Mahalanobis distance, a rotation invariant distance measurement, is calculated 206 between each projected pixel and the centroid of each cluster in the color code book.

[0081] The Mahalanobis distance, r, is calculated according to:

[0082] r²=(x−m_x)^TC_x⁻¹(x-m_x)

[0083] Where x is the pixel vector, m_xand C_xare the mean and covariance of each code book cluster, respectively.

[0084] Pixels whose Mahalanobis distance to a particular centroid is within a specified threshold are tentatively assigned 208 a label from the code book corresponding to that particular centroid's surface color class. Preferably, a normalized threshold of 4 to 10 provides sufficient discrimination between possible color classes.

[0085] Regions of pixels are then formed 210 which consist of adjacent pixels having the same color class label.

[0086] For each such region formed, all the pixels in that region are reprojected 212 into the color space of the color code book to form a test cluster. The orientation of this test cluster, determined by calculating the eigen vector of its covariance matrix 214, is then compared 216, with the orientation of all the clusters of the tentatively assigned color surface class from the color code book.

[0087] In practice, the cluster orientations, Ψ_ij, within a class in a color code book vary from α to β and a test cluster's orientation which falls between α−5° and β+5° is considered to match a class from the color code book.

[0088] The region whose test cluster passes the orientation comparison is determined to truly be of that particular color surface class that was tentatively assigned earlier. Failure of the orientation test determines that this particular region is not a representation of the expected color surface class.

[0089] This method first assures a match of a region to some appearance of a surface class within the code book and the later verification ensures a spectral match between the region and at least one cluster of that particular surface class. If a pixel is assigned to multiple color surface classes, an image smoothing filter 218 will eliminate such plurality of assignments so that the end result is a clear detection of the color surface classes present in an image. Finally, to complete region localization and recognition, each detected region within an image is given 220 a region identifier and assigned 220 a color surface semantic label from the color code book.

[0090] The use of the color code book and the detection of color surface classes without pre-segmenting an image have tremendous application to improving database indexing and retrieving of images in response to a color-related query. FIG. 3 illustrates a method of indexing and querying an image database.

[0091] To begin with, each image in a database is analyzed 302. And for every color surface class in the color code book, regions of the images are detected which belong to that color surface class. Preferably, the detection and classification of color surface regions are performed according to the method described in FIG. 2. However, other region localization and detection methods are considered within the scope of the present invention.

[0092] A color region index is then formed 304 containing all regions belonging to a specific colored surface class:

[0093] C_i=>{<R_j, I_{k> . . . }}

[0094] Where Ciis a semantic label for a color surface class and R_jis the region identifier of a region belonging to image I_k.

[0095] By well-known look-up methods, the database receiving 306 a semantic user query for images having sky, grass, skin, etc. will use the color region index to locate 308 and retrieve 310 relevant images and regions that possess the query surface. An appropriate GUI is then used to display the retrieved images for the user 312.

[0096] One powerful aspect of image databases is the possibility of querying for images which are similar to a known image. Thus, a query consisting of an image, rather that a semantic label, is also considered within the scope of an embodiment of the present invention and will result in the retrieval of images comprising similar color surface classes. A method for handling this type of query is illustrated in FIG. 4.

[0097] According to one embodiment of the present invention, in order to perform a query based on an image, known color surface classes within the image are first detected 404 and labeled according to the color code book; preferably using the method described in FIG. 2, although other segmentation methods are considered to be within the scope of the present invention. Next, the semantic labels from the detected, known color surface classes are extracted and used to search 406 the color region index; which was formed earlier 304, in order to locate images containing similar color surface classes.

[0098] All matching images and regions are located 408, retrieved 410 and then presented 412 to the user by an appropriate GUI.

[0099] In the above algorithm, the color code book with its semantic labels allows for the formation of the color region index for the entire database. Performing the indexing by pre-computation avoids searching for matching regions within images during image retrieval. Also, this algorithm allows the retrieving of images based on semantic color queries such as sky, grass, skin, etc. Another benefit derived from this system is that a single image is capable of having multiple semantic color labels and that the indexing method retains these multiple color labels when classifying the images.

[0100] A system and method has been shown in the above embodiments for an image database indexing and querying method which uses a pre-computed color region index to allow content-based color querying of the database by semantic labels or by images. While various preferred embodiments have been shown and described, it will be understood that there is no intent to limit the invention by such disclosure, but rather, it is intended to cover all modifications and alternate constructions falling within the spirit and scope of the invention as defined in the appended claims. For example, the present invention should not be limited by graphical user interface, computer operating system, computer hardware platform, image capturing hardware, initial color-space representation, database management system, or image encoding algorithm. In addition, the present invention can be implemented locally on a single PC, connected workstations (i.e. networked-LAN), across extended networks such as the Internet or using portable equipment such as laptop computers or wireless equipment (RF, microwaves, infrared, photonic, etc.)