JP2001266161A - Method for describing uniform texture and its descriptor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method for effectively and efficiently interpreting and retrieving a uniform texture which can be used in a wide application range for inquiry to a multimedia retrieving engine, an electronic library or a medical picture and in the retrieval of pictures obtained by a remote sensor. SOLUTION: A texture descriptor is provided with a multiple resolution smoothing coherence parameter being the combination of a series of smoothing coherence parameters of the scales of different original pictures. The smoothing coherence parameters are numbers for describing the smoothness of a texture. On the basis of this smoothing coherence parameter, the texture is described by using a picture histogram, the parameter of a deformation GMRF model and features in a conversion domain.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to image retrieval, video editing, medical image inquiry, digital
It can be used to describe homogeneous textures for libraries (electronic libraries), especially for Internet multimedia information retrieval.

[0002]

Background of the Invention Texture is an important feature in the analysis of various types of images, including natural scenery, data acquired by remote sensors, and biomedical aspects. Texture identification is believed to play an important role in the human visual system for recognition and interpretation.

[0003] In recent years, as a result of a worldwide effort to digitize vast records of images, films and videos,
Created a demand for an automatic search system. Tools that support retrieval from texture-rich images have a wide range of uses in video editing, medical image queries, electronic libraries, or the merchandise market.

[0004] Texture features are an essential part of such image retrieval systems. In order to effectively and efficiently represent textures, good quality texture descriptors are required.

In general, a texture is analyzed by a texture model. Since textures have statistical and structural properties, texture models have been historically classified into statistical, structural, and statistical / structural (hybrid) types. Statistical approaches focus on the statistical properties of textures. Texture patterns are characterized by either a statistical or random model of the grayscale values of the image pixels. Early methods include Fourier power spectra, autocorrelation functions, spatial gray scale co-occurrence random, and the like. From the 1970s to the 1990s,
Bayesian statistics (G. Cross and AKJain, Markov random
field texture models, IEEE Trans. PAMI, 5 (1): 25-3
9, 1983) based on Markov random fields (MRF: Mar
An enormous amount of literature has been published in the field of texture modeling using kov random fields (R. Chellappa)
and S. Chatterjee, Classification of Textures Using
Gaussian Markov Random Fields, IEEE Trans. On Sco
ust. Speech Signal Processing, 33: 959-963, 198
Five). Another channel of texture analysis is the multi-channel filtering mechanism (WYMa, Netra: A Too
lbox for Navigating Large Image Databases. PhD the
sis, University of California, Santa Barbara, 199
There is a filtering theory based on 7). In the last decade, weblet filterbanks and Gabor filters have been the mainstream tools for analyzing textures based on statistical features of filtered images.
In the structural approach, textures are represented by texture primitives and their spatial arrangement rules. The main disadvantage of the structural method is that it cannot acquire or generate the randomness inherent to natural textures. Natural textures usually contain statistical and structural components,
A texture model that can express both structure and randomness has been studied. A new model based on the Wald decomposition of a uniform random field has been proposed by Francos et al. (JMFrancos, AZMeiri, and B. Po
rat, A Unified Texture Model Based on a 2-D Wold-l
ike Decomposition. IEEE Trans. Signal Processing,
41: 2665-2678, 1993).

[0006] UCSB and Samsung (P. Wu, WY Ma et
al, Texture Descriptor, ISO / MPEG, P77, Lancaster, 1
999) proposed using a perceptual browsing component (PBC) and a similarity extraction component (SRC) to describe the texture. The PBC is a high-level texture description indicating whether the texture is regular or not, and the main direction and scale of the texture. SRC, on the other hand, describes low-level texture features. Each texture is broken down into 24 subbands by Gabor weblets at different scales and directions. The first and second moments (mean and variance) are stored as sub-band features.

[0007] The ICU and ETR have proposed applying Radon transforms and atoms to pursue texture matching (YMRo, SY. Kim et al, Texture Descri).
ption Using Atoms of Matching Pursuit, ISO / MPEG P6
12, Lancaster, 1999). The Radon transform is used because it is believed to fit the human visual system (HVS). First, a two-dimensional image is converted into a one-dimensional signal (sinogram) using Radon transform. Next, the texture can be represented as an atom (texture descriptor) by performing the matching in the Radon space. Although the document uses only the energy of the atom as a feature, it is assumed that the deviation of the atom, the order of the matched atoms, and the order of importance for the components of the texture descriptor can all be used as features. .

[0008] HHI has proposed using a weblet to filter textures and extract features in the spatial frequency domain (JR Ohm and F. Bunjamin, Descri).
ptorfor Texture in Wavelet Domain, ISO / MPEG, P566,
Lancaster, 1999). In general, any weblet can be used for disassembly. The document uses a three-stage weblet decomposition. To keep the scale and direction unchanged, along each direction and scale, six combined channels, each with three weblet channels, are applied to the descriptor. After threshold quantization for the scale channel, the lower central moments are analyzed. To get the rotation and scale invariant criterion (total 64)
A histogram of the energy, mean, and variance of the block is created. A histogram is also created for the Weblet channel. Matching is performed by comparing the two histograms.

Hyundai proposed a translation, scale, and rotation invariant texture descriptor for Zernike moments for DFT images (DG Sim and HK Kim,
Translation, Scale, and Rotation Invariant Textur
e Descriptor of Zernike Moments on DFT Image, ISO /
MPEG, M4723, Vancouver, 1999). By transforming the input texture image by the DFT process, an absolute DCT coefficient image that is invariant to translation is obtained. The change in scale in the original spatial domain remains in the DCT coefficient image. The DFT coefficient image is normalized by the total AC to obtain scale invariance. Rotation-invariant Zernike moments are extracted as texture descriptors, and since DFT preprocessing and scale normalization have been performed, the descriptors are also translation- and scale-invariant.

[0010] These texture descriptors extract different features to describe a uniform texture. But,
Neither of these descriptors attempts to describe the details of the texture. The search results are clearly improved if the information hidden in neighboring pixels is known.

[0011]

According to the inventors' knowledge, there are a number of smooth patches that are either not textured or do not contribute to the main texture of the image, even in a uniform texture. This is a downside when performing texture description in an image. Extracting texture features based on these patches only results in a set of features that cannot represent the main texture in the image. However, smooth patches are an important component that is essential for other textures.

Therefore, the present disclosure intends to describe a texture image only by its main texture, which is not dealt with in the prior art. The problem to be solved is how to remove unnecessary patches in the image in order to improve the texture description ability.

At the same time, depending on the nature of the texture containing both statistical and structural information, the proposed texture descriptor must describe both the statistical and structural properties of the texture. No. Therefore, the second problem to be solved by the present disclosure is how to select a feature for expressing statistical information and structural information included in a texture together.

[0014]

According to the inventors' findings, there are a number of smooth patches that do not contribute to the main texture, especially the texture description, even for very uniform textures. Other textures often have defects that negatively affect texture identification. However, for some textures, the smooth patches contribute significantly to the main texture. These findings indicate that certain texture smooth patches should be removed before extracting texture features, while certain other textures should preserve smooth patches for texture description. Understand.

In order to solve the problem of how to determine whether or not to remove a smooth patch for extracting a feature in a texture image, it is named “Smooth Coherence Parameter (SCP)”. Parameters are proposed in this disclosure. If the smoothing patches of the texture image are clustered together, it is likely that these patches are not part of the main texture in the image. In this case, the smooth coherence parameter should be large. On the other hand, if the smoothing patches are dispersed in the image, these patches are likely to be important components of the main texture in the image. In this case, the smooth coherence parameter should be small. The present invention provides a method for calculating this smooth coherence parameter.
The smooth coherence parameter is a number that describes the smoothness of the texture, but a combination of a series of smooth coherence parameters at different scales of the original image may be used as the multi-resolution smooth coherence parameter.

In order to describe the statistical and structural properties of the texture, it is proposed to use different features to describe these properties. Modified multi-resolution Gaussian
Markov random field (MRGMRF: modified multi
The relationship between adjacent pixels is described using resolution Gaussian Marlovrandom fields, and the spatial distribution of textures that cannot be represented by MRGMRF parameters is described using histograms of main textures. Features in the transform domain are used to describe structural properties.
In order to fuse features, it is important to weight similarity measures.

The present invention provides a method of texture interpretation that can be used for image retrieval (eg, query). In order to accurately extract uniform texture features, we propose a concept of smooth coherence parameter (SCP) based on our findings. Given a threshold, it can be determined whether the smooth portion of the image is part of the texture. Other texture features, Gauss-Markov random field parameters and image histograms are extracted based on the values of the SCP. To clarify the degree of texture structuring,
Features in the transform domain are extracted. Such a uniform texture descriptor consists of smooth coherence parameters, multi-resolution Gaussian Markov random field parameters, image histograms and features from the transform domain. These SCPs are used to describe whether smooth portions of the image are clustered together or dispersed within the image. Gauss-Markov random field parameters are used to describe the relationship between neighboring pixels. Image histograms are applied to describe the distribution of tones in an image that cannot be represented in a Gauss-Markov random field. Features from the transform domain are used to determine the degree of texture structuring. Together with the obtained texture descriptor, the Euclidean distance can be used as a measure of similarity. To improve search accuracy,
Features should be weighted. The descriptor of the image to be searched can be compared with the descriptor stored in the image database. The smaller the distance, the more similar the two textures are. This texture descriptor can be used for still image search, video search, medical image query, and electronic library.

[0018]

DETAILED DESCRIPTION OF THE INVENTION The present invention provides a method for extracting features for describing a uniform texture.

First, a smooth coherence parameter (SCP) is calculated. In such a situation, the smooth coherence parameter is obtained as follows. Define a neighborhood system of the image (eg, a secondary neighborhood system shown in FIG. 2). Scan the image and detect pixels whose neighboring pixels have the same gray scale as themselves. Count the total number of pixels surrounding the neighborhood having the same gray shade as the selected pixel. Calculate the average of the above total number of pixels having the same gray shade. -Divide the average value by the number of pixels surrounding the neighborhood of a certain pixel.

This procedure is represented by the following equation.

(Equation 1)Here, Ω = {(s1, s2): 0 ≦ s1 ≦ M−1, 0 ≦ s2 ≦ N−1} is 2
S = (s1, s2) is the coordinates of the grid points on the two-dimensional grid,
η_sIs near s, W_sIs the smooth coherence parameter c
A predetermined local around site s to calculate
・ Window, N _ηIs the number of pixels in the neighborhood, N_WIs W_s
Is the number of pixels in. If c is 1, huge in image
Indicates that there is a smooth region, where c is
0 indicates that there is almost no smooth area in the image
You. Figure 1 shows two textures and their corresponding smoothing
Shows the coherence parameter. Figure 2 shows the smoothness of the image.
Local window for calculating
1 shows a sequence of a window. Sites labeled '0'
Is a site that is considered. Labeled '1'
Sites are labeled '0' in the second-order neighborhood
In the vicinity of the site. The label labeled '2'
Site surrounds the site labeled '0'
Site. FIG. 3 shows the above-mentioned smooth coherence path.
4 shows a flowchart for calculating parameters.

The GMRF model is effective in expressing a fine texture, but has a weak point in describing a macro texture in which the relationship with a small neighborhood cannot be accurately expressed. Due to limitations in small neighborhoods, the GMRF model cannot be a useful tool for describing general textures. To overcome these shortcomings of the conventional GMRF model, efforts were made in this disclosure to improve texture description capabilities.

To describe a macro texture, a multiresolution representation is applied. The original image is filtered by a series of low-pass filters, downsampled and filtered image sequence _Gl (l =
0,1 ,,,, L-1) is obtained. GMRF is applied to each of these filtered images. This method allows the texture structure represented by the information of two pixels separated from each other in the original image to be adjacent at a higher level in the filtered image. The method can extend the neighborhood of a conventional GMRF model to reveal more information. The parameters of this model can be used as features for image retrieval. These parameters describe the relationship between neighboring pixels at different levels.

The basic notation used for GMRF is as follows: A set of lattice points included in the vicinity of the site s is represented by η _s . X is a two-dimensional lattice Ω = ｛(s1, s2):
Represents a random field on 0 ≦ s1 ≦ M−1, 0 ≦ s2 ≦ N−1}. X _s
Represents a random variable at site s. If X is modeled by a GMRF with symmetric neighborhood η,
X is

(Equation 2) Or

(Equation 3) Can be expressed as Here, B _θ is an M × N matrix, and s = (s ₁ ,
s ₂ ) is the coordinates of the grid point on the two-dimensional grid, and η _s is near s. e is zero mean Gaussian noise, and its autocorrelation is given by:

(Equation 4) Therefore, the GMRF is completely characterized by the set of parameters {θ, σ ² }.

The parameters of the GMRF model are matched least squares (R. Chellappa. Progress in Pattern Recognitio
n, Chapter Two-dimensional Discrete Gaussian Marko
v Random Field Models for Image Processing, pages
79-112. Amsterdam: North-Holland, 1985). When considering a symmetric neighborhood set, θ _r and σ can be evaluated by the following equations.

(Equation 5) and

(Equation 6) Here, q _s represents Col. [X _{s + r} , r∈η _s ] and Col. [X _{s + r} , r∈
eta _s] column vector component consisting _{X s + r, r∈η s,} x s is X
represents the realized value of _s . The modified multi-resolution GMRF is the original G
Based on MRF model. The parameters of the modified multi-resolution GMRF are the corresponding smooth coherence parameters in the filtered image and a predetermined threshold T
Is evaluated according to If c <T, the smooth patch in the image is likely to be part of the main texture, so GM
RF parameters are evaluated based on the pixels of the entire image. Otherwise, the GMRF parameter is a pixel whose neighborhood is not smooth

(Equation 7) Will be evaluated based solely on:

In addition, the histogram of the original image is employed as part of the feature in this descriptor. This histogram can describe the distribution of gray levels in an image that cannot be represented by the parameters of the multi-resolution GMRF model. The histogram of the image x is given by the following equation.

(Equation 8) Where δ (.) Is the Dirac function, Q (.) Is the quantization function,
q is the bin number of the histogram. The histogram of the original image can also be constructed based on the threshold T.
If c <T, a histogram of the entire image is calculated,
Otherwise, only histograms of pixels whose neighborhood is not smooth are calculated. FIG. 4 shows the procedure for calculating the GMRF parameters and the histogram. Where SCP
(X) is a function for calculating a smooth coherence parameter, T is a threshold for determining whether or not a smooth portion in an image is a part of a texture, and modified_GMRF (x) is a parameter for calculating a modified GMRF. Function, normal_GMRF
(X) is a function to calculate standard GMRF parameters, modif
ied_hist (X) is a function for calculating a modified histogram, and normal_hist (X) is a function for calculating a standard histogram.

[0026] In the above description, the characteristics of the texture may be constituted by _{^{(θ r l, σ l,}} H x, c l). Where l
= 0,. . . , L-1 is the low-pass filtering level, L
Is the number of filtered images.

Unfortunately, our experiments have shown that some uniform textures, such as the texture of FIG. 1 (b), cannot be well described by these features. Several features in the transform domain are used to describe textures that are highly structured but cannot be represented by the features described above.

Consider a case where F (.) Is a linear transform such as a fast Fourier transform (FFT) or a discrete cosine transform (DCT), and x is an image, y = F (x). Two methods are proposed to extract features in the transform domain.

The first method is as follows. 1. Data y satisfying y = F (x) is obtained in the transform domain. 2. Discard the DC component of the coefficients in the transform domain. 3. Calculate the magnitude of the coefficients in the transform domain to find the maximum. 4. Scan the magnitude of the coefficients in the transform domain and if the coefficients are greater than a predetermined percentage of the maximum, set the site of the coefficients to '1' in a binary matrix of the same size as the image, otherwise In this case, the site of the coefficient is set to '0' in a binary matrix. 5. Divide the binary matrix into non-overlapping blocks and use the normalized sum of '1' in the selected block as a feature in the transform domain.

The second method is as follows. 1. Data y satisfying y = F (x) is obtained in the transform domain. 2. Discard the DC component of the coefficients in the transform domain. 3. Calculate the magnitude of the coefficients in the transform domain. 4. The matrix y is divided into non-overlapping _blocks , and the normalized sum of the sizes (t ₁ ,..., T _blocks ) in the selected block is used as a feature in the transform domain.

The non-overlapping blocks hatched in FIG. 5 are blocks selected to extract features from the transform domain. The block in the upper left corner is the low frequency block.

The inventors have found that smooth coherence
The parameters are not only parameters suitable for measuring a smooth patch in a texture image, but are also good features for expressing a texture.

Therefore, the descriptor in the present disclosure is represented by the following vector format.

Equation 9] ^{_{F = {θ 1 0, ...}} , θ m 0, θ 1 1, ..., θ m n-1, σ 0, ..., σ n-1, c 0, .. ., c ⁿ⁻¹ , h ₁ , ..., h _bins , t ₁ , ..., t _blocks ｝ (9) where m is a parameter number in the GMRF model,
n is the number of filtered images, bins is a histogram
The number of bins, blocks, is the number of features from the transform domain.

The low-pass filter used to obtain the multi-resolution representation of the image is a low-pass filter of a wavelet. In our experiments, we used the Haar Weblet to simplify the calculations. This makes the low-pass filter very simple.

(Equation 10) That is, the average of four adjacent pixels at level l-1 is used as the value at the corresponding site at level 1 and L = 3. A second order neighborhood system is applied to the GMRF model. Thus, the five parameters (θ ₁ ^l , θ ₂ ^l ,
θ ₃ ^l , θ ₄ ^l , σ ^l ) are extracted for each filtered image x ^l (x ⁰ is the original image x). The original image is quantized to eight gray levels, so that there are eight histogram bins. As shown in FIG. 2, eight features are selected from transform domains (T (1)... T (8)) from 16 blocks. Thus, the descriptor has 15 GMRF parameters,
Consisting of three smooth coherence parameters, eight histogram bins, and eight features extracted from the transform domain by Method 1 or Method 2 as described above, for a total of 34 parameters. The parameters should be normalized because their ranges are different. In this context, normalization is performed by dividing the parameter by the maximum value of the parameter having the same characteristics. For example, the GMRF parameter should be divided by the maximum value of the GMRF parameter. Subsequently, all these features should be quantized to an integer between 0 and 255.

A typical method for feature extraction can be described as follows. 1. The texture image is filtered using a low-pass filter to obtain a series of images. 2. These images are quantized into a predetermined number of gray levels. 3. For each filtered image, scan the image to detect pixels where all neighboring pixels have the same gray shade, and surround the neighborhood of pixels with the same gray shade as the selected pixel Calculate the average of the number of 4. Divide the average by the number of pixels surrounding the neighborhood of a pixel to obtain a smooth coherence parameter for the filtered image. 5. The smooth coherence parameters are calculated for all filtered images according to the above procedure to obtain the multi-resolution smooth coherence parameters. 6. For each filtered image, if the corresponding smooth coherence parameter is less than a predetermined threshold, evaluate the parameters of the Gauss-Markov random field model for all pixels in the image; otherwise, Evaluate the parameters of the Gauss-Markov random field model only for those pixels where not all of the neighboring pixels have the same gray shade as yourself, and use the procedure described above for Gauss-Markov Calculate the parameters of the random field to obtain the modified multi-resolution Gauss-Markov random field model parameters. 7. If the smooth coherence parameter is less than a predetermined threshold, calculate a histogram of the entire image; otherwise, generate a histogram only for those pixels whose neighbors do not all have the same gray shade. calculate. 8. The quantized original image is transformed into a transform domain by FFT or DCT, the DC component of the coefficients in the transform domain is discarded, the magnitude of the coefficients in the transform domain is calculated, and the maximum value is found. 9. Scan the magnitude of the coefficient in the transform domain, and if the coefficient is greater than a predetermined percentage of the maximum, set the site of that coefficient to '1' in a binary matrix of the same size as the image, otherwise In this case, the coefficient site is set to '0' in the binary matrix, the binary matrix is divided into non-overlapping blocks, and the sum of the selected blocks is used as a feature in the transform domain. 9 '. The matrix containing the magnitudes of the coefficients is divided into non-overlapping blocks, and the sum of the magnitudes in the selected blocks is calculated as a feature of the transform domain. 10. Normalize and quantize the parameters to form a texture descriptor.

In the above procedure, procedure 9 and procedure 9 'can be alternatively selected to obtain features from the transform domain. FIG. 6 shows the entire procedure for feature extraction.

As shown in FIG. 6, the original image X is decomposed into a plurality of grayscale level images (multi-resolution decomposition), and after each image is quantized, the SCP is calculated. Histograms and parameters for the multi-resolution GMRF are calculated based on those SCPs. Encoding and generation of a texture descriptor are performed based on these values and the feature amount in the transform domain.

The syntax for a texture descriptor is given as follows. texture_descriptor {for (i = 0; i <layers; i ++) {for (j = 0; j <directions; j ++) theta_i_j 8 uimsbf sigma_i 8 uimsbf c_i 8 uimsbf} for (i = 0; i <bins; i ++) H_i 8 uimsbf for (i = 0; i <blocks; i ++) T_i 8 uimsbf}

The meaning of the descriptor is very simple. theta_
i_j is the j-th GMRF parameter of the i-th filtered and down-sampled image. sigma_i
Is the GMRF standard deviation of the ith filtered and downsampled image. c_i is the smooth coherence parameter of the ith filtered and downsampled image. H_i is the ith bin of the histogram of the original image. levels is the number of images filtered and downsampled. directions is the number of parameters of the GMRF model. bin is the number of histogram bins. blocks is the number of blocks selected in the transform domain.

The parameters of the modified GMRF are used to interpret the relationship between adjacent pixels. Smooth coherence parameters are used to describe the distribution of smooth patches in an image. The histogram of the original image is used to describe the distribution of gray levels in the image. Parameters from the transform domain are used to describe the structuredness of the texture. All of these features constitute a compact and effective texture descriptor. A uniform texture can be effectively and efficiently interpreted by the above texture descriptor.

Together with the obtained texture descriptor, the Euclidean distance can be used as a measure of similarity. Features should be weighted to improve search accuracy. The distance between the query image (Q) and the test image (T) from the image database is measured by:

[Equation 11] Here, w _i , i = 1,. . . , 34 are feature weights. In this similarity measure, the smaller the distance between the two images, the more similar the two images. Then the distance between the query descriptor and the descriptor stored in the database is sorted. The user can obtain a similar texture from the image database using the query texture. For a 34-dimensional texture descriptor according to the invention,
The weights are given as follows.

(Equation 12)

[0042]

According to the present invention, a uniform texture can be described very effectively and efficiently. That is, by using the smooth coherence parameter, the search speed in the image search becomes extremely high. With the present invention, a new generation of multimedia
Provided are texture descriptors useful for search engines, electronic libraries and image retrieval.

[Brief description of the drawings]

FIG. 1 shows two textures (a, b) and their smooth coherence parameters.

FIG. 2 is a diagram showing a neighborhood and a local window used in the present invention.

FIG. 3 is a flowchart for calculating a smooth coherence parameter (SCP) (SCP (“0”,
The local window to calculate the "1", "2" sites (all sites) and the secondary neighborhood of the "0" site (sites labeled "1").

FIG. 4 is a flowchart for calculating a GMRF parameter and a histogram.

FIG. 5 is a diagram illustrating blocks selected for feature extraction in a transform domain.

FIG. 6 is a diagram showing a procedure of feature extraction.

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor One Ray Singapore 534415 Singapore, Thai Sen Avenue, Block 1022, 04-3530, Thai Sen Industrial Estate, Panasonic Singapore Institute F Terms (reference) 5L096 FA14 FA23 FA35

Claims (36)

[Claims] 1. A multi-resolution smooth coherence parameter which is a combination of a series of smooth coherence parameters of different scales of the original image, wherein said smooth coherence parameter is a number for describing the smoothness of the texture. Texture descriptor to describe the texture to be featured. 2. A texture descriptor for describing a texture, comprising a parameter of a modified multi-resolution Gaussian Markov random field. 3. A texture descriptor for describing a texture, characterized by comprising image histogram parameters. 4. A texture descriptor for describing a texture, comprising a feature of an image in a transform domain. 5. A texture descriptor for describing a texture, comprising a multi-resolution smooth coherence parameter and a modified multi-resolution Gaussian Markov random field parameter. 6. A texture descriptor for describing a texture, comprising a multi-resolution smooth coherence parameter and an image histogram parameter. 7. A texture descriptor for describing a texture, comprising a multi-resolution smooth coherence parameter and features of an image in a transform domain. 8. A texture descriptor for describing a texture, comprising a multi-resolution smooth coherence parameter, a modified multi-resolution Gaussian Markov random field parameter, and an image histogram parameter. 9. A texture descriptor for describing a texture, comprising a multi-resolution smooth coherence parameter, an image histogram parameter, and features of an image in a transform domain. 10. A texture for describing a texture, comprising multi-resolution smooth coherence parameters, modified multi-resolution Gaussian Markov random field parameters, and image features in a transform domain. descriptor. 11. A texture comprising a multi-resolution smooth coherence parameter, a modified multi-resolution Gaussian Markov random field parameter, an image histogram parameter, and image features in the transform domain. Texture descriptor to describe the. 12. A texture description method, comprising: calculating a multi-resolution smooth coherence parameter; and encoding the parameter into an encoded representation. 13. A texture description method, comprising: estimating parameters of a modified multiresolution Gaussian Markov random field; and encoding the parameters into an encoded representation. 14. A method for describing a texture, comprising: calculating image histogram parameters; and encoding the parameters into an encoded representation. 15. A method for describing a texture, comprising: calculating features of an image in a transform domain; and coding the parameters into a coded representation. 16. Computing a multi-resolution smooth coherence parameter, evaluating a parameter of a modified multi-resolution Gaussian Markov random field, and encoding the parameter into an encoded representation. A texture description method, characterized in that: 17. A method comprising calculating a multi-resolution smoothing coherence parameter, calculating an image histogram parameter, and encoding the parameter into an encoded representation. How to describe the texture. 18. The method of claim 18 further comprising the steps of: calculating a multi-resolution smooth coherence parameter; calculating image features in the transform domain; and encoding the parameter into an encoded representation. The description method of the texture. 19. calculating a multi-resolution smooth coherence parameter; evaluating a parameter of a modified multi-resolution Gaussian Markov random field; calculating an image histogram parameter; Encoding a texture into a textured expression. 20. A step of calculating a multi-resolution smoothing coherence parameter, a step of calculating an image histogram parameter, a step of calculating characteristics of an image in the transform domain, and encoding the parameter into an encoded representation. A texture description method. 21. calculating a multi-resolution smooth coherence parameter; evaluating a parameter of a modified multi-resolution Gaussian Markov random field; calculating features of the image in a transform domain; Encoding the texture into an encoded representation. 22. A step of calculating a multi-resolution smooth coherence parameter; evaluating a parameter of a modified multi-resolution Gaussian Markov random field; calculating an image histogram parameter; A method for describing a texture, comprising: calculating a feature; and encoding the parameter into an encoded representation. 23. The method according to claim 12, wherein the multi-resolution smoothing coherence parameter is calculated using a low-pass filter. Filtering the images in the set of images, quantizing each image in the set of images to a plurality of gray levels, and calculating a smooth coherence parameter for each image in the set of images. Performing the steps of: 24. Claims 13, 16, and 1
9. A method for estimating parameters of the modified multi-resolution Gaussian Markov random field model according to any one of claims 21 and 22, wherein the texture image is converted into a series of images using a low-pass filter. Filtering to a group of images, quantizing each image in the series of images to a plurality of gray levels, and calculating a smooth coherence parameter for each image in the series of images; Calculating Gauss-Markov random field parameters for all filtered images to obtain modified multi-resolution Gauss-Markov random field model parameters. 25. Claims 14, 17, and 1
9. A method for calculating an image histogram according to any one of claims 20 and 22, wherein the texture image is quantized into a predetermined number of gray shades, and the image comprises a smooth coherence image. Calculating a parameter;
Calculating an image histogram over the entire image if the parameter is less than a predetermined threshold;
If the smooth coherence parameter is greater than or equal to the threshold value, calculating a histogram only for pixels where not all of the adjacent pixels have the same gray shade as the pixel itself. how to. 26. Claims 15, 18, and 20.
23. A method for calculating features in a transform domain according to any one of claims 22 to 22, wherein the step of quantizing the texture image to a predetermined number of gray shades comprises: Converting to a coefficient, discarding the DC component of the coefficient in the transform domain, calculating the magnitude of the residual coefficient in the transform domain, finding a maximum value among the magnitudes of the coefficient And a step of defining a binary matrix having the same size as the image, wherein each component of the binary matrix is set to '1' if a coefficient at a corresponding position is larger than a predetermined ratio of the maximum value. Setting, otherwise setting to '0', and using the sum of '1' and '0' terms in each row and column of the binary matrix as a feature in the transform domain thing Wherein. 27. Claim 15, Claim 18, and Claim 2.
23. A method according to any one of claims 21 and 22, wherein the method comprises the steps of: quantizing a texture image into a predetermined number of gray shades; Transforming the resulting image into coefficients in a transform domain, discarding the DC component of the coefficients in the transform domain, calculating the magnitude of the remaining coefficients in the transform domain, and the magnitude of the coefficients Finding a maximum value in the matrix; defining a binary matrix having the same size as the image; and determining that the coefficient at the corresponding position is larger than a predetermined ratio of the maximum value. Setting each component to '1'; otherwise, to '0'; dividing the binary matrix into non-overlapping blocks; transforming the sum of the selected blocks Method characterized by comprising the step of using a feature of the in. 28. A method for calculating the smooth coherence parameter according to any one of claims 23 to 25, further comprising defining a neighborhood around each of the selected pixels. Scanning the image to find pixels having the same gray shade as all of the neighboring pixels; and counting the total number of pixels surrounding the neighborhood having the same gray shade as each of the selected pixels. Calculating the average of said total number of pixels having the same gray shade, and dividing said average by the number of pixels surrounding the neighborhood of a pixel. 29. The method of calculating a Gauss-Markov random field parameter according to claim 24, further comprising: if the smooth coherence parameter is smaller than a predetermined threshold, all the pixels in the image. Estimating the parameters of the Gauss-Markov random field model over pixels of the pixel, and if the smooth coherence parameter is equal to or greater than a predetermined threshold, then all pixels in the neighborhood will have the same gray shade as themselves. Gauss-Markov only for pixels that do not have
Evaluating the parameters of the random field model. 30. The method according to claim 26, wherein the transform domain is a discrete cosine transform domain. 31. The method according to claim 26, wherein the transform domain is a Fourier transform domain.
7. The method according to 7. 32. The method according to claim 23, wherein the low-pass filter is a low-pass weblet filter. 33. The encoding method according to claim 12, wherein a parameter of the descriptor is normalized to [0, 1], and the parameter is set to 0 to 255. Quantizing to an integer in the range 34. Obtaining a texture descriptor for a query image; calculating a distance between the texture descriptor and each texture descriptor retrieved in a database of texture descriptors; And selecting a texture having a smaller distance as a matched texture. 35. Obtaining a texture descriptor for a query image; calculating a Euclidean distance between the texture descriptor and each retrieved texture descriptor in a database of texture descriptors; Sorting in ascending order and selecting a texture having a smaller distance as a matched texture. 36. Obtaining a texture descriptor of a query image; calculating a weighted Euclidean distance between said texture descriptor and each retrieved texture descriptor in a database of texture descriptors; Sorting the distances in ascending order and selecting a texture having a smaller distance as a matched texture.