HK1060634B - Image retrieval using distance measure - Google Patents

Description

Background

This invention relates to image processing and, more particularly, to the characterization of images.

Image processing refers to the analysis and manipulation of images using a computer. The discipline encompasses a wide array of techniques, including pattern recognition and other image analysis, compression and encoding, such as image transmission, and image construction, to name but a few examples. Image processing is used in diverse fields such as astronomy, medicine, military operations, communications, geology, meteorology and so on.

Although visual in nature, images may be characterized so that a computer or other processor-based system may also "see" the image or, for example, distinguish the image from other images. Identifying the moment, entropy, center of mass, orientation, or other histogram-based features is one approach to image characterization. Structural and syntactic features of the image may serve as characteristic parameters. Geometric indicia, such as perimeter, area, eccentricity, or Euler number, may numerically identify the image.

Because images are often perturbed, such as during transmission, the characteristic parameters of the image are ideally invariant following transformations. Image transformations may include translation, rotation, scaling, shearing, compression, inclusion of noise, and so on.

Particularly in the emerging domain of Internet technology, image searching may be an important tool. Different characteristics of the images, used for image searching, may have widely varying results, however. This type of image searching is known as content-based image retrieval.

One technique for retrieving images from a database of images is to associate each image with a certain number of features. Using feature vectors, each image, represented by its feature values, is mapped to a point in an n-dimensional feature space, where there are n features identified. A similarity query is then used to retrieve the images lying in a neighborhood close to the query image.

In executing these similarity queries, the precision of obtaining the image may be offset by the size of the image database and the time to calculate the various features involved in the similarity query.

Kato et al. describe in "A handwritten character recognition system using directional element feature and asymmetric Mahalanobis distance", IEEE transactions on pattern analysis and machine intelligence, IEEE Inc. New York, US, vol. 21, no.3, 1999-03, pages 258-262, a recognition system for chinese or japanese characters. A directional feature vector of a query character image is calculated by extracting information about its contours and dot orientation and compared with the directional feature vectors of character images stored in a database. First, a rough classification using the city block distance is performed, which singles out a few candidates from the great number of possible characters. Afterwards, a fine classification based on an asymmetric Mahalanobis distance is carried out.

Further, Kato et al. disclose in "A handwritten character recognition system using modified Mahalanobis distance", Systems&computers in Japan, Scripta technical journals, New York, US, vol.28, no.1, 1997, pages 46-54, a system similar to that described in the last paragraph in an earlier stage of development. Comparative measurements are performed to find out the best distance function for character recognition. For the chosen circumstances, the Mahalanobis distance proves to be the best evaluation function.

Kapoor et al. provide in "New techniques for exact and approximate dynamic closest-point problems", Proceedings of the tenth annual symposium on computational geometry, Stony Brook, NY, USA, 6-8 June 1994, pages 165-174, a theoretical basis for new techniques for solving closest-point problems using range trees.

Flickner et al. propose in "Query by image and video content: The QBIC system" Computer, IEEE Computer Society, Long Beach, CA, US, vol.28, no. 9, 1995-09-01, pages 23-32, a system which allows image queries on large databases. The images are described by certain characteristic features derived from colours, textures, shapes etc. which are stored in a database. The query image features can be compared with the stored features of the example images by means of distance functions.

Niblack et al. explain the last-mentioned system in "The QBIC project: Qerying images by content using color, texture, and shape", Proceedings of the Spie, Spie, Bellingham, VA, US, vol. 1908, 1993, pages 173-187, in more detail and identifies the distance functions used for comparing the images. Depending on the chosen features, either the Euclidean distance or edge matching algorithms are applied.

De Maesschalck et al. examine and interpret the Mahalanobis distance in relation to the Euclidean distance in "The Mahalanobis distance", Chemometrics and intelligent laboratory systems, Elsevier Science Publishers, Amsterdam, NL, vol.50, 2000-01, pages 1-18.

US-A-6104833 again discloses a character recognition apparatus. A query pattern and different prototype characters are each characterised by two predetermined features. In some cases, the Euler number of an image may be used to determine if an image is a candidate to match a query. The distance between the feature vectors is calculated, for example using the city block distance function, and it is therefrom decided which character is represented by the query pattern.

Dey et al. provide in "A fast algorithm for computing the Euler number of an image and its VLSI implementation", Thirteenth international conference on VLSI design, IEEE, 2000-01-03, to 2000-01-07, pages 330-335, an algorithm for computing the Euler number of a binary image.

US-A-5642431 describes a system for detecting human faces in images. It includes a processor which determines a distance metric between a preprocessed query image and different prototype images. The distance metric may contain two components, one based on a Mahalanobis distance, the other based on a Euclidean distance.

Thus, there is a continuing need for optimizing image retrieval methods.

The invention solves this problem by a method, system and computer program according to the independent claims. The dependent claims refer to preferred embodiments of the invention.

Brief Description of the Drawings

• Figure 1 is a block diagram of a binary image according to one embodiment of the invention;
• Figure 2 is a block diagram of a 4- and 8-neighborhood according to one embodiment of the invention;
• Figure 3 is a block diagram of a gray-tone image according to one embodiment of the invention;
• Figures 4A through 4D are binary images of the gray-tone image of Figure 3 according to one embodiment of the invention;
• Figures 5A through 5D are reflected gray code representations of the gray-tone image of Figure 3 according to one embodiment of the invention;
• Figure 6 is a block diagram of a system according to one embodiment of the invention;
• Figure 7 is a flow diagram for calculating a variance-covariance matrix for plural images of the system of Figure 6 according to one embodiment of the invention; and
• Figure 8 is a flow diagram for calculating the Mahalanobis distance between a query image and plural images of the system of Figure 6 according to one embodiment of the invention.

Detailed Description Euler Number

To retrieve an image from a database of images, each image may be characterized to distinguish it from the other images. For example, the image may be characterized numerically.

For example, each binary image of a database of binary images may be assigned an Euler number. An Euler number of a binary image is the difference between the number of connected components (objects) in the image and the number of holes in the image.

The Euler number is a topological feature of the binary image. The Euler number remains invariant under translation, rotation, scaling and rubber-sheet transformation of the binary image.

A binary image 16 may be represented in an N × M pixel matrix 20, in which an object pixel 12 is shaded, to indicate a binary value of 1, while a background pixel 12 is unshaded, to indicate a binary value of 0, as depicted in Figure 1. Within the N × M pixel matrix 20, a connected component is a set of object pixels such that any object pixel 12 in the set is in the eight (or four) neighborhood of at least one object pixel 12 of the same set. Both a 4-neighborhood 14 and an 8-neighborhood 18 are depicted in Figure 2.

In addition to connected components in the binary image, in one embodiment, a hole is defined as a set of background pixels, such that any background pixel in the set is in the 4 (or eight) neighborhood of at least one background pixel of the same set and this entire set of background pixels is enclosed by a connected component.

A run of the i^th column (or row) of the pixel matrix 20 is defined to be a maximal sequence of consecutive ones in the i^th column (or row). Two runs appearing in two adjacent columns (rows) each are said to be neighboring if at least one pixel of a run is in the eight (or four) neighborhood of a pixel of the other run.

If the binary image 16, known as I, consists of a single row (column) i, the Euler number E(I) of the image 16 is the same as the number of runs, denoted as R(i), of the image 16, as shown by the following formula: $E\left(I\right)=R\left(i\right)$ However, where the binary image 16 consists of multiple rows (columns), the binary image, I, may be divided into partial images, I_i. Further, the Euler number of the entire image can be found using the following relations: $E\left(I\right)=E\left(I_1,\cup ,I_2\right)=E\left(I_1\right)+E\left(I_2\right)-E\left(I_1,\cap ,I_2\right),$ and $E\left(I_{i-1},\cap ,I_i\right)=O_i$ where O_i equals the number of neighboring runs between two consecutive rows (columns), e.g., the (i-1)^th and the i^th rows (columns).

According to one embodiment, the Euler number of a binary image is defined as the difference between the sum of the number of runs for all rows (columns) and the sum of the neighboring runs between all consecutive pairs of rows (columns) of the N x M pixel matrix. Stated mathematically: $E\left(I_N\right)=\sum _{i=1}^{N}R\left(i\right)-\sum _{i=2}^N{O}_i$ where I_N=I denotes the entire image. Thus, by knowing the number of runs and the number of neighboring runs in the binary image, I, the Euler number, E(I) may be computed.

Euler Vector

The Euler number for binary images may be extended to gray-tone images. For example, as depicted in Figure 3, a gray-tone image 36 may be represented in an N × M matrix in which each element 22 is an integer between 0 and 255 decimal denoting the intensity of the corresponding pixel. The intensity value of each element may be represented by an 8-bit binary vector (b₇, b₆, b₅, b₄, b₃, b₂, b₁, b₀). The gray-tone image 36 thus comprises eight bit planes, or eight distinct binary images.

The hypothetical gray-tone image 36 includes three 8-bit pixels 22a, 22b, and 22c, each of which represents a distinct intensity value, as shown.

In one embodiment, the first four most significant bit planes (b₇, b₆, b₅, and b₄) of the 8-bit binary vector are retained. In essence, four binary images are retained. In Figures 4A-4D, the four binary images 18a-18d are depicted, representing the four most significant bit planes of the gray-tone image 36 of Figure 3. An Euler number may be computed for each binary image 18a-18d.

In one embodiment, before computing the Euler number for each binary image 18, the four-bit binary vector (b₇, b₆, b₅, b₄), is converted to its corresponding reflected gray code vector (g₇, g₆, g₅, g₄), in which:

• g7=b7
• g6 = b7 ⊕ b6
• g5 = b6 ⊕ b5
• g4 = b5 ⊕ b4

where ⊕ denotes XOR (modulo-2) operation. The result of the reflected gray code conversion is depicted in binary images 28a through 28d of Figures 5A-5D, respectively, according to one embodiment. It is well-known that for any binary vector, the corresponding reflected gray code is unique and vice-versa.

Accordingly, in one embodiment, the Euler vector, ξ_K of the original gray-tone image, K, is a 4-tuple (E₇, E₆, E₅, E₄), where E_i is the Euler number of the partial binary image, g_i, formed by the i^th bit plane, 4 ≤ i ≤ 7, corresponding to the reflected gray code representation of intensity values.

Thus, in addition to calculating an Euler number for a binary image, an Euler vector for a gray-tone image may be calculated by considering the four most significant binary bit planes of the image, converting the 4-bit binary vector to its corresponding reflected gray code, and computing the Euler number for each reflected gray code representation.

In characterizing gray-tone images using Euler vectors, the elements of the Euler vector have, in some cases, a decreasing order of importance. That is, element E₇ is most significant in characterizing the image, next is E₆, and so on. Further, the ranges and the variances of various elements of the Euler vector widely differ. In such an environment, the Mahalanobis distance may be adopted to provide a measure that captures the similarity and dissimilarity of properties for features of a set of images.

Mahalanobis Distance

The Euler number E_i, may thus characterize a binary image. By extension, an Euler vector ξ_i, including four Euler numbers, E₇, E₆, E₅, and E₄, may characterize a gray-tone image. Each Euler number E_i, may be described as a "feature" of the gray-tone image, according to one embodiment. In one embodiment, Euler vector characterization may form a basis for image retrieval.

Suppose an Euler vector, ξ_K, describes features for each gray-tone image, K, in an image database. A new image, q, having features defined by Euler vector ξ_q, is to be retrieved from the image database. The newly received image, q, may be compared to each of the images in the image database. Once a match is found, image retrieval is complete.

One way to compare two things is to measure a distance between them. Distance measurement may be performed between scalars, such as numbers, but also may be performed between vectors. Where an image is described using a vector, the distance between the vector and a second image having its own vector may be computed. Accordingly, the distance between the two images may be computed.

Suppose the query image 30q is not found in the image database 40. Rather than comparing for an exact match, in one embodiment, a distance measure between the query image 30q and other images in the database may be taken. One distance measure is known as a Mahalanobis distance measure. The Mahalanobis distance measure calculates the "distance" between the query image 30q and the other images in the database, then ranks the distance results. The smallest distance result is the "match."

The Mahalanobis distance measure operation is depicted in Figure 6, according to one embodiment. A system 100 includes a processor 80, an image database 40 and a Mahalanobis distance measure 60. The image database 40 includes multiple images 30, in which each has an associated Euler vector, ξ_K. The Mahalanobis distance measure 60 receives each of the images 30 from the image database 40, and compares them to a query image 30q, which includes its own Euler vector, ξ_q. In one embodiment, the Mahalanobis distance measure 60 is a software program, executable by the processor 80, for performing Mahalanobis distance measurements between the query image 30q and one or more images 30 from the image database 40.

A result table 50, according to one embodiment, includes distance calculations as between the query image 30q and one or more images 30 in the image database 40. In one embodiment, the distance results are ranked according to their size. Thus, the first (or last) entry in the result table 50 is the "match" of the distance calculation.

Each of the images 30 in the image database 40 may be classified or described according to the Euler vector, ξ_K. The elements of the Euler vector, or Euler numbers, E₇, E₆, E₅, and E₄, represent features of the gray-tone image 30 they classify. Conventionally, distance measures between two vectors may be used to capture the "Closeness" of the two vectors. For example, the Euclidean distance between vectors x and y, where x has four elements, x₁, x₂, x₃, and x₄ and y likewise has four elements, y₁, y₂, y₃, and y₄ is given by the following equation: $d\left(x,,y\right)=\sqrt{\left((x_1-y_1{)}^2+(x_2-y_2{)}^2,+,(,x_3,-,y_3,,{)}^2,+,(,x_4,-,y_4,,{)}^2\right)}$

For some phenomena, the Euclidean distance does not provide a very good distance measure. Recall that when characterizing gray-tone images by an Euler vector, the elements of the vector have a decreasing level of importance. That is, E₇ is more important in characterizing the gray-tone image 30 than E₆; E₆ is more important in characterizing the image 30 than E₅; and so on.

Further, the ranges and the variances of the elements (E₇, E₆, E₅, and E₄) of the vector widely differ in their magnitudes. In such an environment, the Mahalanobis distance may provide a better measure than the Euclidean distance for capturing the similarity/dissimilarity properties between the images 30 of the database 40.

The Mahalanobis distance between two vectors, x and y, may be derived from the following equation: $d\left(x,,y\right)=\sqrt{\left(\left(x,-,y\right),,ʹ,*,{M_{\mathrm{VC}}}^{-1},*,\left(x,-,y\right)\right)}$ where ' represents matrix transposition. M_VC is the variance-covariance matrix formed by the image feature vectors: x' = (x₁, x₂, ..., x_N) and y'= (y₁, y₂, ..., y_N). Here, (x - y)' is a (1 × n) matrix, M_VC ^-1 is an (n x n) matrix, and (x - y) is an (n × 1) matrix; hence their product is a scalar.

The variance-covariance matrix of x, denoted as M_VC, is an N × N matrix given by: $\begin{array}{c}{M}_{\mathrm{VC}}\left(i,,j\right)={\mathrm{variance\; of\; the\; i}}^{\mathrm{th}}{\mathrm{feature\; x}}_i,\mathrm{if\; i}=j;\\ ={\mathrm{covariance\; of\; the\; i}}^{\mathrm{th}}{\mathrm{variable\; x}}_i{\mathrm{and\; jth\; variable\; x}}_j\mathrm{if\; i}\ne j\end{array}$ Note that M_VC(i, j) = M_VC(j, i). The variance-covariance matrix of x is also known as a dispersion matrix.

To compute the variance-covariance matrix of image feature vectors, both the variance and the covariance are computed. The variance of a feature, x_i, is defined as: $\mathrm{Variance}\left(x_i\right)={{\delta }_{\mathrm{xi}}}^2=\left(\sum _{k=1}^L,(x_{\mathrm{ik}}-\mathrm{mean}\left(x_i\right){)}^2\right)/L$ where L is the number of the observations of the variable (feature) x_i.

The features of a gray-tone image may also be correlated. The measure of the correlation between any two features is known as covariance. Covariance is defined as follows: $\mathrm{Covariance}\left(x_i,,x_j\right)=\mathrm{cov}\left(x_i,,y_j\right)=\left(\sum _{k=1}^L,\left(x_{\mathrm{ik}},-,\mathrm{mean},\left(x_i\right)\right),,\left(x_{\mathrm{ij}},-,\mathrm{mean},\left(x_j\right)\right)\right)/L$

Once the variance and covariance values are determined, the variance-covariance matrix, M_VC, may be computed. The variance-covariance matrix, M_VC, looks like: $M_{\mathrm{VC}}=\left(\begin{array}{ccccc}{{\delta }_{x1}}^2& \mathrm{cov}\left(x_2,,x_1\right)& \mathrm{cov}\left(x_3,,x_1\right)& \dots & \mathrm{cov}\left(x_n,,x_1\right)\\ \mathrm{cov}\left(x_1,,x_2\right)& {{\delta }_{x2}}^2& \mathrm{cov}\left(x_3,,x_2\right)& \dots & \mathrm{cov}\left(x_n,,x_3\right)\\ \mathrm{cov}\left(x_1,,x_3\right)& \mathrm{cov}\left(x_2,,x_3\right)& {{\delta }_{x3}}^2& \dots & \mathrm{cov}\left(x_n,,x_3\right)\\ ⋮\\ \mathrm{cov}\left(x_1,,x_{\left(n,-,1\right)}\right)& \mathrm{cov}\left(x_2,,x_{\left(n,-,1\right)}\right)& \mathrm{cov}\left(x_3,,x_{\left(n,-,1\right)}\right)& \dots & \mathrm{cov}\left(x_n,,x_{\left(n,-,1\right)}\right)\\ \mathrm{cov}\left(x_1,,x_n\right)& \mathrm{cov}\left(x_2,,x_n\right)& \mathrm{cov}\left(x_3,,x_n\right)& \dots & {{\delta }_{\mathrm{xn}}}^2\end{array}\right)$

In Figure 7, the variance-covariance matrix, M_VC, for a database of images 30, is computed according to one embodiment. To illustrate, assume the image database 40 includes images 30 in which each image 30_K has an associated Euler vector, ξ_K. Each Euler vector, ξ_K includes Euler numbers E_K7, E_K6, E_K5, and E_K4, which constitute "features" of the image 30_K.

The variance-covariance matrix may be computed for all the images 30 of the database 40. Subsequently, a query image 30q may be received by the system 100, as depicted in Figure 6. Once the variance-covariance matrix, M_VC, is calculated, image queries may quickly retrieved desired images.

In one embodiment, the Mahalanobis distance measure 60 computes the mean (or average) of each feature of the image database 40 (block 202). Because the Euler vector, ξ_K, includes four features, Euler numbers E₇-E₄, four mean calculations are computed, as follows: ${\mu }_{E7}=\left(E_{a7},+,E_{b7},+,E_{c7},+,\dots ,+,E_{j7}\right)/j$ ${\mu }_{E6}=\left(E_{a6},+,E_{b6},+,E_{c6},+,\dots ,+,E_{j6}\right)/j$ ${\mu }_{E5}=\left(E_{a5},+,E_{b5},+,E_{c5},+,\dots ,+,E_{j5}\right)/j$ ${\mu }_{E4}=\left(E_{a4},+,E_{b4},+,E_{c4},+,\dots ,+,E_{j4}\right)/j$ where j is the number of images in the image database 40.

The features of each image 30 in the database 40 may be correlated. The measure of the correlation between any two features is reflected in the covariance equation, defined above. Accordingly, in one embodiment, the covariance between features of the images 30 is calculated (block 206). For example, in the illustrated database 40, the following covariance equations may be computed: $\mathrm{cov}\left(E_7,,E_6\right)=\mathrm{cov}\left(E_6,,E_7\right)=\left(\left(E_{a7},-,{\mu }_{E7},\left),\right(,E_{a6},-,{\mu }_{E6},),+,(,E_{b7},-,{\mu }_{E7},\left),\right(,E_{b6},-,{\mu }_{E6}\right),+,(,E_{c7},-,{\mu }_{E7},),(,E_{c6},-,{\mu }_{E6},),+,\dots ,+,(,E_{j7},-,{\mu }_{E7},),(,E_{j6},-,{\mu }_{E6},),/,j\right)$ $\mathrm{cov}\left(E_7,,E_5\right)=\mathrm{cov}\left(E_5,,E_7\right)=\left(\left(E_{a7},-,{\mu }_{E7},\left),\right(,E_{a5},-,{\mu }_{E5},),+,(,E_{b7},-,{\mu }_{E7},\left),\right(,E_{b5},-,{\mu }_{E5}\right),+,(,E_{c7},-,{\mu }_{E7},),(,E_{c5},-,{\mu }_{E5},),+,\dots ,+,(,E_{j7},-,{\mu }_{E7},),(,E_{j5},-,{\mu }_{E5},),/,j\right)$ $\mathrm{cov}\left(E_7,,E_4\right)=\mathrm{cov}\left(E_4,,E_7\right)=\left(\left(E_{a7},-,{\mu }_{E7},\left),\right(,E_{a4},-,{\mu }_{E4},),+,(,E_{b7},-,{\mu }_{E7},\left),\right(,E_{b4},-,{\mu }_{E4}\right),+,(,E_{c7},-,{\mu }_{E7},),(,E_{c4},-,{\mu }_{E4},),+,\dots ,+,(,E_{j7},-,{\mu }_{E7},),(,E_{j4},-,{\mu }_{E4},),/,j\right)$ $\mathrm{cov}\left(E_6,,E_5\right)=\mathrm{cov}\left(E_5,,E_6\right)=\left(\left(E_{a6},-,{\mu }_{E6},\left),\right(,E_{a5},-,{\mu }_{E5},),+,(,E_{b6},-,{\mu }_{E6},\left),\right(,E_{b5},-,{\mu }_{E5}\right),+,(,E_{c6},-,{\mu }_{E6},),(,E_{c5},-,{\mu }_{E5},),+,\dots ,+,(,E_{j6},-,{\mu }_{E6},),(,E_{j5},-,{\mu }_{E5},),/,j\right)$ $\mathrm{cov}\left(E_6,,E_4\right)=\mathrm{cov}\left(E_4,,E_6\right)=\left(\left(E_{a6},-,{\mu }_{E6},\left),\right(,E_{a4},-,{\mu }_{E4},),+,(,E_{b6},-,{\mu }_{E6},\left),\right(,E_{b4},-,{\mu }_{E4}\right),+,(,E_{c6},-,{\mu }_{E6},),(,E_{c4},-,{\mu }_{E4},),+,\dots ,+,(,E_{j6},-,{\mu }_{E6},),(,E_{j4},-,{\mu }_{E4},),/,j\right)$ $\mathrm{cov}\left(E_5,,E_4\right)=\mathrm{cov}\left(E_4,,E_5\right)=\left(\left(E_{a5},-,{\mu }_{E5},\left),\right(,E_{a4},-,{\mu }_{E4},),+,(,E_{b5},-,{\mu }_{E5},\left),\right(,E_{b4},-,{\mu }_{E4}\right),+,(,E_{c5},-,{\mu }_{E5},),(,E_{c4},-,{\mu }_{E4},),+,\dots ,+,(,E_{j5},-,{\mu }_{E5},),(,E_{j4},-,{\mu }_{E4},),/,j\right)$ The covariance is computed as between Euler numbers E₇ and E₆, E₇ and E₅, E₇ and E₄, E₆ and E₅, E₆ and E₄, and E₅ and E₄. The covariance is associative. Thus, cov(E₇, E₆)=cov(E₆, E₇), and so on.

Now that the variance and covariance are known for the features E7, E6, E5, and E4 of the image database 40, the variance-covariance matrix M_VC may be computed according to one embodiment (block 208). For a database of images characterized using Euler vectors, ξ_K, the variance-covariance matrix M_vc looks like: $M_{\mathrm{VC}}=\left|\begin{array}{cccc}{{\delta }_{E7}}^2& \mathrm{cov}\left(E_6,,E_7\right)& \mathrm{cov}\left(E_5,,E_7\right)& \mathrm{cov}\left(E_4,,E_7\right)\\ \mathrm{cov}\left(E_7,,E_6\right)& {{\delta }_{E6}}^2& \mathrm{cov}\left(E_5,,E_6\right)& \mathrm{cov}\left(E_4,,E_6\right)\\ \mathrm{cov}\left(E_7,,E_5\right)& \mathrm{cov}\left(E_6,,E_5\right)& {{\delta }_{E5}}^2& \mathrm{cov}\left(E_4,,E_5\right)\\ \mathrm{cov}\left(E_7,,E_4\right)& \mathrm{cov}\left(E_6,,E_4\right)& \mathrm{cov}\left(E_5,,E_4\right)& {{\delta }_{E4}}^2\end{array}\right|$ Computation of the variance-covariance matrix, according to Figure 7, is thus complete.

Once the variance-covariance matrix, M_VC, is known, the Mahalanobis distance may be measured between a query image 30q and one or more images 30 in the image database 40.

In one embodiment, the Mahalanobis distance is computed according to the flow diagram of Figure 8 (block 302). First, a difference between two vectors is calculated. Each vector includes feature information for a given image 30. For example, where the Mahalanobis distance is sought between the query image 30q and the image 30a of the image database 40, the difference between the Euler vectors ξ_q and ξ_a is calculated. The difference between ξ_q and ξ_a is: $\left({\xi }_q,-,{\xi }_a\right)=\left(\begin{array}{c}\left(E_{q7},-,E_{a7}\right)\\ (E_{q6}-E_{a6})\\ (E_{q5}-E_{a5})\\ (E_{q4}-E_{a4})\end{array}\right)$ Likewise, the transpose (ξ_q - ξ_a)' is as follows: $({\xi }_q-{\xi }_a)ʹ=\left(\begin{array}{c}\left(E_{q7},-,E_{a7}\right)\end{array},\begin{array}{c}\left(E_{q6},-,E_{a6}\right)\end{array},\begin{array}{c}\left(E_{q5},-,E_{a5}\right)\end{array},\begin{array}{c}\left(E_{q4},-,E_{a4}\right)\end{array}\right)$

Additionally, to compute the Mahalanobis distance between image 30q and any of the images 30 in the image database 40, the inverse of the variance-covariance matrix, M_VC, or M_VC ^-1, is calculated (block 304). The inverse of the matrix M_VC is calculated according to known principles.

From these operations, the Mahalanobis distance may be calculated (block 306). For example, using the query image 30q, the Mahalanobis distance may be calculated between the query image 30q and one or more of the images 30 in the image database 40, as follows: $d\left({\xi }_q,,{\xi }_a\right)=({\xi }_q-{\xi }_a)ʹ*{M_{\mathrm{VC}}}^{-1}*({\xi }_q-{\xi }_a)$ $d\left({\xi }_q,,{\xi }_b\right)=({\xi }_q-{\xi }_b)ʹ*{M_{\mathrm{VC}}}^{-1}*({\xi }_q-{\xi }_b)$ $\begin{array}{c}d\left({\xi }_q,,{\xi }_c\right)=({\xi }_q-{\xi }_c)ʹ*{M_{\mathrm{VC}}}^{-1}*({\xi }_q-{\xi }_c)\\ ⋮\\ d\left({\xi }_q,,{\xi }_j\right)=({\xi }_q-{\xi }_j)ʹ*{M_{\mathrm{VC}}}^{-1}*({\xi }_q-{\xi }_j)\end{array}$

In one embodiment, the Mahalanobis distance measure 60 ranks the distance calculations in order from smallest to largest. Accordingly, the first entry into the result table 50 is the closest match for the query image 30q. In one embodiment, using the Euler vector as the representative image feature and the Mahalanobis distance as the closeness measure, the desired image is found to lie at minimum distance. Noisy and compressed versions of the query image 30q have next higher rank in the result table 50.

In some embodiments, the various features used to characterize the images 30 may not be equally important for retrieval. In such a case the elements of the variance-covariance matrix, M_vc, may be multiplied by a suitable weight factor according to the importance of the corresponding feature.

For very large databases, computing the distance from the query image 30q to all other images 30 in the database 40 may take substantial time. Thus, ranking and, in turn, retrieval, may be undesirably slow. To obviate the problem, multi-dimensional search techniques may be employed to retrieve a subset of images lying in the neighborhood of the query image 30q in the feature space. Ranking according to the proposed distance measure may then be performed on this subset only.

Range Trees

There are many techniques available for finding an item in a database. A sequential search of the database may occur, a hash table may be used and binary search techniques may be employed, as examples. A binary tree is a data structure used for storage, removal, and searching of items in a data set. With respect to the images 30 in the image database 40, it is desirable to find the image 30 that is similar to the query image 30q without searching all j of the images in the database 40.

In one embodiment, a multi-dimensional search tree is thus used for storing the images 30. The multi-dimensional search tree permits efficient retrieval of the images 30.

In one embodiment, images are stored in the binary tree at random. Sometimes, however, the data is not inserted into the binary tree in random order. The performance of a binary search, therefore, may be less than optimal. The binary tree may be "rebalanced" after each insertion or deletion of data. During rebalancing operations, the nodes of the binary tree are rearranged using rotations. Typically, these rotations minimize the height of the binary tree.

One type of binary tree is known as an AVL tree (named for its creators Adelson-Velskii and Landis). In this particular form of binary tree, rebalancing ensures that, for each node in the AVL tree, the difference in height between its subtrees, known as a balance factor, is not greater than 1.

In one embodiment, a four-dimensional range tree 62, as depicted in Figure 6, is used as a data structure for storing the Euler vectors, ξ_K, associated with each image 30 in the database 40. The range tree 62 is four-dimensional because the Euler vector includes four features, Euler numbers E₇, E₆, E₅, and E₄. Using the Euler vector, ξ_q, for the query image 30q, the range tree 62 specifies a particular node. In one embodiment, the Mahalanobis distance measure 60 is performed between the query image 30q and only these images lying at the node, rather than on all images 30 of the range tree 62. The results may be ranked in descending order, such that the first distance value points to the matched image 30.

While the present invention has been described with respect to a limited number of embodiments, those skilled in the art will appreciate numerous modifications and variations therefrom. It is intended that the appended claims cover all such modifications and variations as fall within the scope of this present invention.

Claims (19)

1. A computer-implemented method for similarity-based image retrieval comprising the following steps:

   characterising a plurality of gray-tone images (30) included in a database (40) using Euler vectors which are calculated by considering the most significant binary bit planes of the gray-tone images, wherein the elements of the Euler vectors are Euler numbers;

   receiving a query image (30q) including its own characterising Euler vector;

   computing a Mahalanobis distance between the query image (30q) and said plurality of images (30) from the image database (40) using the characterising Euler vectors; and

   retrieving the image from the database (40) which has the smallest of said computed distances as match for the received query image (30q).
2. The method of claim 1, wherein computing a Mahalanobis distance between the query image (30q) and said plurality of images (30) from the image database (40) further comprises:

   constructing a variance-covariance matrix using the Euler vectors of the plurality of images (30);

   selecting a first image from the plurality of images (30), the first image comprising the Euler vector; and

   computing the Mahalanobis distance between the first image and the query image (30q) using the variance-covariance matrix.
3. The method of claim 2, wherein constructing a variance-covariance matrix using the Euler vectors of the plurality of images (30) further comprises:

   computing the mean of an Euler vector element of the plurality of images (30);

   computing the variance of the Euler vector element of the plurality of images (30); and

   computing the covariance between a pair of Euler vector elements of the plurality of images (30).
4. The method of claim 3, further comprising:

   constructing the variance-covariance matrix from the computed variances and covariances.
5. The method of claim 2, wherein computing the Mahalanobis distance between the first image and the query image using the variance-covariance matrix further comprises:

   calculating the difference between the Euler vector characterising the first image and the Euler vector characterising the query image to produce a difference vector.
6. The method of claim 5, further comprising:

   multiplying the difference vector by an inverse of the variance-covariance matrix to produce an intermediate result; and

   multiplying the intermediate result by an inverse of the difference vector.
7. The method of claim 6, further comprising:

   storing the Euler vectors of the plurality of images (30) in a range tree (62).
8. A system (100) comprising:

   a processor (80);

   an image database (40) including plural gray-tone images (30), each of the images (30) characterised using an Euler vector which is calculated by considering the most significant binary bit planes of the gray-tone image (30), wherein the elements of the Euler vector are Euler numbers; and

   a machine-readable storage (70) medium comprising:

   a software program that if executed by the processor (80), is effective to receive a query image (30q) including its own characterising Euler vector, to compute a Mahalanobis distance between the query image (30q) and said plural images (30) from the database (40) using the characterising Euler vectors, and to retrieve the image from the database (40) which has the smallest of said computed distances as match for the received query image (30q).
9. The system (100) of claim 8, wherein the software program, if executed by the processor (80), is effective to:

   construct a variance-covariance matrix using the Euler vectors of the plural images (30) in the image database (40) .
10. The system (100) of claim 9, wherein the software program further computes the Mahalanobis distance by:

    computing the mean of one or more Euler vector elements of the plural images (30);

    computing the variance of one or more Euler vector elements of the plural images (30); and

    computing the covariance between pairs of Euler vector elements of the plural images (30).
11. The system (100) of one of claims 8 to 10, wherein the software program further computes the Mahalanobis distance by:

    calculating a difference between the Euler vector characterising a first image and the Euler vector characterising a query image to produce a difference vector; and

    calculating an inverse of the difference vector.
12. The system (100) of claim 11, wherein the software program further computes the Mahalanobis distance by:

    multiplying the difference vector and an inverse of the variance-covariance matrix to produce an intermediate result; and

    multiplying the intermediate result by an inverse of the difference vector.
13. A computer program for similarity-based image retrieval storing instructions to enable a processor-based system (100) to:

    characterise a plurality of gray-tone images (30) included in a database (40) using Euler vectors which are calculated by considering the most significant binary bit planes pf the gray-tone images (30), wherein the elements of the Euler vectors are Euler numbers;

    receive a query image (30q) including its own characterising Euler vector;

    compute a Mahalanobis distance between the query image (30q) and said plurality of images (30) from the image database (40) using the characterising Euler vectors; and

    retrieve the image from the database (40) which has the smallest of said computed distances as match for the received query image (30q).
14. Computer program of claim 13, further storing instructions to enable a processor-based system (100) to:

    construct a variance-covariance matrix using the Euler vectors of the plurality of images (30);

    select a first image from the plurality of images (30), the first image comprising the Euler vector; and

    compute the Mahalanobis distance between the first image and the query image (30q) using the variance-covariance matrix.
15. The computer program of claim 14, further storing instructions to enable a processor-based system (100) to:

    compute the mean of an Euler vector element of the plurality of images (30);

    compute the variance of the Euler vector element of the plurality of images (30); and

    compute the covariance between a pair of Euler vector elements of the plurality of images (30).
16. The computer program of claim 15, further storing instructions to enable a processor-based system (100) to:

    construct the variance-covariance matrix from the computed variance and covariance.
17. The computer program of claim 16, further storing instructions to enable a processor-based system (100) to:

    calculate the difference between the Euler vector characterising the first image and the Euler vector characterising the query image (30q) to produce a difference vector.
18. The computer program of claim 17, further storing instructions to enable a processor-based system (100) to:

    multiply the difference vector by an inverse of the variance-covariance matrix to produce an intermediate result; and

    multiply the intermediate result by an inverse of the difference vector.
19. The computer program of claim 18, further storing instructions to enable a processor-based system (100) to:

    store the Euler vectors of the plurality of images (30) in a range tree (62).