JP4394399B2 - Image analysis apparatus, image analysis program, storage medium, and image analysis method - Google Patents

Description

  The present invention relates to an image analysis device, an image analysis program, a storage medium, and an image analysis method.

  A wide range of research has been conducted on area segmentation for texture images. Many of them are based on a statistical approach based on a co-occurrence matrix, Markov Random Field (MRF), Fourier transform, and the like. One of the main purposes of region segmentation based on this kind of approach is to segment a texture image into regions with a high degree of freedom under the evaluation of uniformity.

  When evaluating such uniformity, the handling of the range of pixels on which texture analysis is performed becomes one major focus. More specifically, depending on the purpose of the image processing, the image may be grouped into several large regions using global features rather than dividing the image into several smaller regions based on local texture features. May be desirable. For example, this is a case where an object having a superior texture feature, such as a patterned shirt, is cut out with high accuracy. In such a case, if the analysis accuracy is improved by finely dividing the inside of the area corresponding to the object, the boundary of the area is distorted. That is, there is a problem of a trade-off between improvement in analysis accuracy by expanding the range of pixels on which texture analysis is performed and distortion at the region boundary.

  As a method for dealing with this type of problem, Non-Patent Document 1 proposes a method of making the analysis range variable by estimating the uniform range.

Toshio Uchiyama, Naoki Takegawa, Taichi Nakamura, and Hiroshi Kaneko, “Area Segmentation of Texture Images Using Uniform Range Estimation,” IEICE Theory (D-II), vol. J83-D-II, pp. 1446-1459 2000.

  However, since the method proposed in Non-Patent Document 1 is a method of capturing features based on the second-order MRF with respect to the density value, the processing target is limited to a texture having a relatively short period.

  In recent years, multi-resolution analysis using wavelets has attracted attention as a means for efficiently capturing fluctuation characteristics of density values in various scales in the space-frequency domain. In particular, wavelet frames have been actively taken up as shift-invariant analysis means for texture features. However, in order to capture global texture features that are characterized by various scales, it is necessary to analyze a region corresponding to that range.

  An object of the present invention is to provide an image analysis apparatus, an image analysis program, a storage medium, and an image analysis method capable of performing object extraction with high analysis accuracy and little boundary distortion.

An image analysis apparatus according to the present invention includes an area dividing unit that divides an image into predetermined regions, and a multi-resolution analysis that repeatedly executes a multi-resolution analysis using wavelet frames for each of the regions divided by the region dividing unit. Means, a feature extraction means for extracting a predetermined feature quantity from the wavelet coefficients obtained by the multi-resolution analysis means, and a region label for setting a label for each area based on the feature quantity extracted by the feature extraction means A setting unit; a clustering unit that clusters each region with a predetermined feature amount in a predetermined region as a cluster center; and a boundary between the predetermined region and a region that includes the predetermined cluster obtained by the clustering and a surrounding region Region boundary optimization that optimizes shape according to predetermined criteria It includes a stage, a.

According to the present invention, in the image analysis apparatus, the region dividing unit, the multi-resolution analyzing unit, the feature extracting unit, and the region label setting unit are recursively executed until a predetermined end condition is satisfied.

According to the present invention, in the image analysis apparatus, the region dividing unit, the multi-resolution analyzing unit, the feature extracting unit, and the clustering unit are recursively executed until a predetermined end condition is satisfied.

According to the present invention, in the image analysis device according to any one of the above, the area dividing unit divides the target image into a predetermined number of rectangular areas having a predetermined size.

Further, according to the present invention, in the image analysis device according to any one of the above, the region label setting unit assigns a predetermined label according to a predetermined similarity regarding a feature amount between neighboring regions.

Further, the present invention is the image analysis device according to any one of the above, wherein the region boundary optimization unit introduces a Markov random field model to the region and the surrounding region, and the likelihood of the region based on the model. Search for the boundary with the highest degree.

According to the present invention, in the image analysis device according to any one of the above, the feature extraction unit uses a vector having each element of average energy obtained for each predetermined component from each predetermined decomposition level as a feature amount.

The image analysis program of the present invention is installed in a computer and has an area dividing function for dividing an image into predetermined areas, and multi-resolution analysis using wavelet frames for each of the areas divided by the area dividing function. A multi-resolution analysis function that is repeatedly executed, a feature extraction function that extracts a predetermined feature amount from wavelet coefficients obtained by the multi-resolution analysis function, and each region based on the feature amount extracted by the feature extraction function A region label setting function for setting a label, a clustering function for clustering each region with a predetermined feature amount in a predetermined region as a cluster center, a region composed of the predetermined region and a predetermined cluster obtained by the clustering, and The boundary shape with the surrounding area Executing a region boundary optimization to optimize according to the criteria of the constant, to the computer.

In the image analysis program, the present invention causes the computer to recursively execute the region division function, the multi-resolution analysis function, the feature extraction function, and the region label setting function until a predetermined end condition is satisfied.

In the image analysis program, the present invention causes the computer to recursively execute the region division function, the multi-resolution analysis function, the feature extraction function, and the clustering function until a predetermined end condition is satisfied.

A storage medium of the present invention stores any one of the image analysis programs described above.

In addition, the image analysis method of the present invention includes an area dividing step for dividing an image into predetermined areas, and a multi-resolution that repeatedly executes multi-resolution analysis using wavelet frames for each of the areas divided by the area dividing step to a predetermined level. An analysis step, a feature extraction step for extracting a predetermined feature amount from the wavelet coefficients obtained by the multi-resolution analysis step, and a region in which a label is set for each region based on the feature amount extracted by the feature extraction step A label setting step, a clustering step of clustering each region with a predetermined feature amount in a predetermined region as a cluster center, a region composed of the predetermined region and a predetermined cluster obtained by the clustering, and a surrounding region Boundary shape is given It includes a region boundary optimization step of optimizing accordance quasi, the.

According to the present invention, in the image analysis method, the region dividing step, the multi-resolution analyzing step, the feature extracting step, and the region label setting step are recursively executed until a predetermined end condition is satisfied.

According to the present invention, in the image analysis method, the region dividing step, the multi-resolution analysis step, the feature extraction step, and the clustering step are recursively executed until a predetermined end condition is satisfied.

According to the present invention, in the image analysis method according to any one of the above, the region dividing step divides the target image into a predetermined number of rectangular regions having a predetermined size.

According to the present invention, in the image analysis method according to any one of the above, the region label setting step assigns a predetermined label in accordance with a predetermined degree of similarity related to a feature amount between neighboring regions.

Further, the present invention is the image analysis method according to any one of the above, wherein the region boundary optimization step introduces a Markov random field model to each of the region and the surrounding region, and the likelihood of the region based on the model. Search for the boundary with the highest degree.

  According to the present invention, in the image analysis method according to any one of the above, the feature extraction step uses, as a feature amount, a vector having each element of average energy obtained for each predetermined component from each predetermined decomposition level.

  According to the present invention, object segmentation with high analysis accuracy and low boundary distortion is performed by performing region boundary optimization after rough region division based on space-frequency analysis at various scales using wavelet frames. It can be carried out.

  An embodiment of the present invention will be described with reference to FIGS.

  FIG. 1 is a block diagram schematically showing a hardware configuration of an image analysis apparatus 1 to which the present invention is applied. As shown in FIG. 1, the image analysis apparatus 1 is, for example, a personal computer or a workstation, and includes a CPU (Central Processing Unit) 2 that is a main part of the computer and controls each part centrally. The CPU 2 is connected by a bus 5 to a ROM (Read Only Memory) 3 which is a read-only memory storing BIOS and a RAM (Random Access Memory) 4 which stores various data in a rewritable manner. Since the RAM 4 has a property of storing various data in a rewritable manner, the RAM 4 functions as a work area for the CPU 2 and functions as an input buffer, for example.

  Further, the bus 5 includes an HDD (Hard Disk Drive) 6 that is an external storage device, a CD-ROM drive 8 that reads a CD (Compact Disc) -ROM 7 as a mechanism for reading a computer program that is a distributed program, A communication control device 10 that controls communication between the image analysis device 1 and the network 9, an input device 11 such as a keyboard or a mouse that functions as input means, and a display device such as a CRT (Cathode Ray Tube) or LCD (Liquid Crystal Display). 12 are connected via an I / O (not shown). Various types of information are stored in the HDD 6 in the form of files and folders in the file system structure on the computer.

  A CD-ROM 7 shown in FIG. 1 implements the storage medium of the present invention, and stores an OS (Operating System) such as Windows (registered trademark) and various computer programs. The CPU 2 reads the computer program stored in the CD-ROM 7 with the CD-ROM drive 8 and installs it in the HDD 6.

  As the storage medium, not only the CD-ROM 7 but also various types of media such as semiconductor memory such as various optical disks such as DVD, various magnetic disks such as various magneto-optical disks and flexible disks, and the like can be used. Further, a computer program may be downloaded from the network 9 such as the Internet via the communication control device 10 and installed in the HDD 6. In this case, the storage device that stores the computer program in the transmitting server is also a storage medium of the present invention. Note that the computer program may operate on a predetermined OS (Operating System), and in that case, the OS may take over the execution of some of the various processes described below, It may be included as part of a group of program files that constitute predetermined application software, OS, or the like.

  The CPU 2 that controls the operation of the entire system executes various processes based on a computer program loaded on the HDD 6 used as the main storage of the system.

  Next, the contents of arithmetic processing executed by the CPU 2 of the image analysis apparatus 1 based on the computer program will be described. Here, the characteristic processing of the present embodiment among the various types of arithmetic processing executed by the CPU 2 of the image analysis device 1 will be described below.

  FIG. 2 is a flowchart schematically showing processing executed by the image analysis apparatus 1. As shown in FIG. 2, the CPU 2 of the image analysis apparatus 1 first initializes a counter n to 0 (step S1).

Next, an image to be analyzed (original image) is divided into a predetermined rectangular area (hereinafter referred to as a tile) with a predetermined size (step S2). For example, a method of equally dividing the original image into (2 ^{n + 1} ) ² tiles can be considered.

  Next, wavelet frame decomposition is executed for each tile up to a predetermined decomposition level (step S3). The simplest implementation of wavelet frame decomposition is to omit downsampling in wavelet decomposition. Further, the limit of the decomposition level may be determined in consideration of the size of the target tile.

  Subsequently, feature vectors are extracted for each tile (step S4). By the wavelet frame decomposition in step S3, coefficient values corresponding to the four types of LL component, HL component, LH component, and HH component are obtained for each decomposition level by the same number as the number of pixels of the tile. Here, L is an alphabet representing a low frequency and H is a high frequency. When two of these are arranged, the alphabet positioned in front corresponds to the horizontal direction, and the alphabet positioned behind corresponds to the vertical direction. The average energy of each component may be adopted as the feature vector element. Note that it is not always necessary to employ all decomposition levels and all components as elements of feature vectors. For example, when it is desired to suppress the influence of fine periodic fluctuations, there is a method that does not use a decomposition level component using a frame with a short period. In addition, when the energy of the low frequency component becomes too high, it may be considered not to use the LL component at each decomposition level.

  Next, the counter n is compared with the upper limit M of the number of repetitions (step S5). When the counter n has not reached the upper limit M of the number of repetitions (Y in step S5), region label setting processing for setting the label of each tile based on the feature vector extracted in step S4 is performed (step S6).

  Here, the region label setting process in step S6 will be described. FIG. 3 is a flowchart schematically showing the region label setting process. As shown in FIG. 3, first, a variable k and a variable l representing the position of the tile in the image are each initialized to 0, and 0 is assigned to all the tiles as initial labels (step S21). Then, 0 is assigned to the counter j (step S22).

Next, an error d between the tile (k, l) and the tile (k + τ _j , l + τ _j ) (j = 0, 1,..., J−1) located in the vicinity thereof is calculated (step S23). ). Here, J represents the number of neighboring tiles. As the value of the error d, a square error between feature vectors of each tile may be used. Also, the range of neighboring tiles is arbitrarily set by the designer.

Thereafter, the error d is compared with a predetermined threshold value D (step S24). When the error d has reached the threshold value D (N in step S24), the process proceeds to step S26. On the other hand, when the error d has not reached the threshold value D (Y in step S24), the same label is newly assigned to the tile (k, l) and the neighboring tile (k + τ _j , l + τ _j ) (step S25). . Specifically, for example, the label values may be made to correspond in accordance with the following conditions (1) to (4). Here, the value of the label in the tile (k, l) is represented as s (k, l).

(1) When s (k, l) = 0 and s (k + τ _j , l + τ _j ) = 0, both are given new labels.

(2) When s (k, l) = 0 and s (k + τ _j , l + τ _j ) ≠ 0, s (k, l) ← s (k + τ _j , l + τ _j ) is assumed.

(3) When s (k, l) ≠ 0 and s (k + τ _j , l + τ _j ) = 0, s (k + τ _j , l + τ _j ) ← s (k, l).

(4) When s (k, l) ≠ 0 and s (k + τ _j , l + τ _j ) ≠ 0, the error between s (k + τ _j , l + τ _j ) is not between any other neighboring tiles. When it is lower than the error, s (k, l) ← s (k + τ _j , l + τ _j ) is set.

In the case of (4), the label value can be changed in accordance with the change of s (k, l) for other tiles having the same value as s (k, l) before the label change. Conceivable. Specifically, among such tiles, a tile having the same value as s (k + τ _j , l + τ _j ) exists in the vicinity, and an error with the tile is lower than an error with any other neighboring tiles. For example, a method of recursively changing labels.

  In step S26, the counter j is compared with the number J of neighboring tiles. If the counter j has not reached the number J of neighboring tiles (Y in step S26), the counter j is incremented by 1 (step S27), and the processes in and after step S23 are repeated. On the other hand, when the counter j has reached the number J of neighboring tiles (N in step S26), the process proceeds to step S28.

  In step S28, the variable k is compared with the number of tiles K in the horizontal direction. When the variable k has not reached the number of tiles K in the horizontal direction (Y in step S28), the variable k is incremented by 1 (step S29), and the processes after step S22 are repeated. On the other hand, when the variable k has reached the number of tiles K in the horizontal direction (N in step S28), the process proceeds to step S30.

  In step S30, the variable l is compared with the number of tiles L in the vertical direction. When the variable l has not reached the number of tiles L in the vertical direction (Y in step S30), 0 is substituted for the variable k and the variable l is incremented by 1 (step S31), and the processes in and after step S22 are repeated. On the other hand, when the variable l has reached the number of tiles L in the vertical direction (N in step S30), the region label setting process ends.

  When the area label setting process (step S6) is thus completed, the counter n is incremented by 1 (step S7), and the processes after step S2 are repeated.

  On the other hand, when the counter n reaches the upper limit M of the number of repetitions (N in step S5), the process proceeds to step S8.

  In step S8, the difference between the counter n and the upper limit M of the number of repetitions is determined. When the counter n is equal to the upper limit M of the number of repetitions (Y in step S8), a predetermined area (hereinafter referred to as a selection area) is selected (step S9), and the process proceeds to clustering after step S11. For the selected area, an area label that matches the area to be cut out, such as an object, may be selected. For example, for the purpose of extracting an object with a texture feature that is known to exist in the vicinity of the center of gravity of the image in advance, a method of selecting a label in a predetermined area close to the coordinates corresponding to the center of gravity of the image is considered. It is done.

  On the other hand, when the upper limit M of the number of repetitions is different from the counter n (N in step S8), the process proceeds to step S10. In step S10, the counter n is compared with the upper limit N of the number of repetitions. When the counter n has not reached the upper limit N of the number of repetitions (Y in step S10), the cluster center is set from the feature vector in the selected area (step S11). The cluster center may be set based on the k-means algorithm, the LBG algorithm, or the like. Further, the number of cluster centers is arbitrary.

  Subsequently, a clustering process is performed on all tiles (step S12), the counter n is incremented by 1 (step S7), and the processes after step S2 are repeated.

  Here, the clustering process in step S12 will be described. FIG. 4 is a flowchart schematically showing clustering processing using the subspace method. For the clustering process, another method such as an LBG algorithm may be used. As shown in FIG. 4, first, a variable c representing a cluster in the selected region, a variable k representing a tile arrangement, and a variable l are each initialized to 0 (step S41).

  Next, the eigenvalue and eigenvector are obtained for the cluster c of the selected region, and a subspace is defined by the eigenvector corresponding to the predetermined higher eigenvalue (step S42).

  Thereafter, based on the partial space, the degree of matching r between the feature vector of the tile (k, l) and the cluster c is measured (step S43). Specifically, the inner product of the eigenvector and the feature vector constituting the partial space may be obtained. Note that when the tile (k, l) is included in the selection region, the process of step S43 may be skipped.

  After calculating the fitness r, the value is compared with a predetermined threshold R (step S44). When the fitness r has reached the threshold value R (Y in step S44), it is determined that the tile (k, l) is a tile having high similarity to the selected region (hereinafter referred to as a similar region candidate tile). (Step S45), the process proceeds to Step S46. On the other hand, when the fitness r does not reach the threshold value R (N in step S44), the process directly proceeds to step S46.

  In step S46, the variable k is compared with the number of tiles K in the horizontal direction. As a result of the comparison between the variable k and the horizontal tile number K, if the variable k has not reached the horizontal tile number K (Y in step S46), the variable k is incremented by 1 (step S47). The processes after S42 are repeated. On the other hand, when the variable k has reached the number of tiles K in the horizontal direction (N in step S46), the process proceeds to step S48.

  In step S48, the variable l is compared with the number of tiles L in the vertical direction. If the variable l has not reached the number of tiles L in the vertical direction (Y in step S48), 0 is substituted for the variable k and the variable l is incremented by 1 (step S49), and the processes in and after step S42 are repeated. When the variable l has reached the number of tiles L in the vertical direction (N in step S48), the process proceeds to step S50.

  In step S50, the variable c is compared with the upper limit C of the number of clusters. The upper limit C of the number of clusters is predetermined by the designer. When the variable c has not reached the upper limit C of the number of clusters (Y in step S50), the variable c is incremented by 1 (step S51), and the processes after step S42 are repeated.

  On the other hand, when the variable c has reached the upper limit C of the number of clusters (N in step S50), a set of extracted similar region candidate tiles that are directly or indirectly adjacent to the selected region is set as the similar region. (Step S52). Here, being indirectly adjacent means that the selected area can be reached through a tile that is a similar area. The process of step S52 is for excluding tiles located at positions distant from the image space from the similar region in order to preserve the spatial continuity of the region, and the similar region candidate tile adjacent to the selected region tile. Can be easily executed by recursively following.

  As described above, in the clustering process, a partial space is configured in the selected area based on the cluster center set in step S11, and the degree of matching r with each partial space is measured for all tiles. , Those having a fitness r value exceeding a threshold value are defined as similar regions. However, tiles that are not adjacent to the selected area are excluded. Note that a tile that can reach a selection area through a tile that is a similar area is considered to be adjacent to the selection area.

  Thus, the clustering process in step S12 ends.

  On the other hand, when the counter n has reached the upper limit N of the number of repetitions (N in step S10), the process proceeds to step S13. In step S13, a region composed of the selected region and a predetermined cluster obtained from the result of the clustering process in step S12 (hereinafter referred to as a similar region, and the selected region and the similar region are collectively referred to as a central region) and its surroundings The boundary with a predetermined area (hereinafter referred to as a peripheral area) is deformed and optimized as necessary.

Here, the region boundary optimization process using the likelihood function in step S13 will be described. FIG. 5 is a flowchart schematically showing region boundary optimization processing using a likelihood function. The likelihood function is calculated using a Gaussian Markov random field model (Gaussian MRF (GMRF) model). As shown in FIG. 5, first, variables corresponding to pixels (u _i , v _i ) ((u _i , v _i ) ∈ center region, i = 0, 1,..., Mc−1) in the center region. e (u _i , v _i ) is 1, and pixel (s _j , t _j ) ((s _j , t _j ) ∈ central region, j = 0, 1,..., Ma−1 in a predetermined peripheral region The variable e (s _j , t _j ) corresponding to) is initialized to −1 (step S61). Note that the peripheral area may be constituted by an arbitrary tile that is directly or indirectly adjacent to the central area.

Next, the variable e ′ (x _h , y _h ) is set to e (x _h , y _h ) (e (x _h , y _h ) = 1 or e (x _h , y _h ) = − 1, h = 0, 1,..., Mc + Ma−1), and the likelihood of the region where e (x _h , y _h ) = 1 is p1, and the region where e (x _h , y _h ) = − 1 Is assigned to p−1. Further, 0 is assigned to each of the variable m and the variable w. (Step S62).

Thereafter, whether or not the pixel (x _m , y _m ) corresponds to the boundary between the central region and the peripheral region is examined based on the value of e (x _m , y _m ) (step S63). Specifically, when e (x _m , y _m ) is different from neighboring e (x _m + δ _i , y _m + δ _i ) (i = 0, 1,..., I), the pixel (x _m , y _m ) may be determined as the boundary. Here, I represents the number of neighboring pixels. As for neighboring pixels, it is only necessary to consider 4 neighborhoods or 8 neighborhoods.

When it is determined that the pixel _(x m, _{y m)} and a boundary (Y in step S63), e substituting _(x m, _{y m) '(x} m, _{y m)} -e to' (step S64), Proceed to step S65. On the other hand, when it is determined that the pixel (x _m , y _m ) is not a boundary (N in step S63), the process proceeds to step S69.

In step S65, the likelihood of the region where e ′ (x _h , y _h ) = 1 is set to p ′ ₁ , and the likelihood of the region where e ′ (x _h , y _h ) = − 1 is set to p ′ _−1. Respectively.

Next, the value of p ′ ₁ p ′ ₋₁ is compared with the value of p ₁ p ₋₁ (step S66). p When _'1 p' _-1 is greater than _p 1 _{p -1} is 1 to w with substitutes (Y in step S66), _{e (x} m, _{y m)} -e to _(x m, _{y m)} Substitute (step S67) and proceed to step S68. On the other hand, when p ′ ₁ p ′ ₋₁ is equal to or less than p ₁ p ₋₁ (N in step S66), the process proceeds to step S68 as it is.

In step S68, substitutes _{(x m, y m) '} -e to _{(x m, y m)'} e.

  In step S69, the variable m is compared with the sum of the number of pixels Mc in the central region and the number of pixels Ma in the peripheral region. When the variable m does not reach Mc + Ma (Y in step S69), the variable m is incremented by 1 (step S70), and the processes after step S63 are repeated. On the other hand, when the variable m has reached Mc + Ma (N in step S69), the process proceeds to step S71.

  In step S71, the difference between the variable w and the variable l is determined. When the variable w is equal to the variable l (Y in step S71), the processes after step S62 are repeated. On the other hand, when the variable w is not the variable l (N in step s71), the process is terminated.

  In this way, in the space-frequency domain, a Gaussian Markov random field model (Gaussian MRF (GMRF) model) is applied to each cluster in the central region and each cluster in the peripheral region, and labels for boundary pixels in the central region are optimized. To do. In other words, after considering the wavelet coefficient of the image as a stochastic process, the probability based on each model parameter is obtained for each boundary pixel, and the region label corresponding to the model with the maximum is repeatedly performed, thereby performing region processing. Perform boundary optimization.

  Here, FIG. 6 is an explanatory diagram showing an example of extracting a texture of interest using a wavelet frame. Here, it is assumed that a shirt object is cut out from the sample image in FIG. For the selected area, the label of the area closest to the coordinates corresponding to the center of gravity of the image was selected. The average energy of each component in the space-frequency domain was adopted as the feature vector element. The number of cluster centers at the time of clustering was set to 3 and determined by the k-means method. When the region boundary was optimized, the entire central region and the entire peripheral region were each treated as one cluster. FIG. 6B shows a cut out region. The area shown in FIG. 6B not only covers the object of interest as a whole, but also has less boundary distortion. However, although it is a part, there is a boundary where the shirt is cut. In this regard, improvement can be expected by newly providing constraints on the shape of the boundary, a model, and the like.

  As described above, in the present embodiment, the region boundary is optimized after rough region division based on space-frequency analysis at various scales using wavelet frames. This makes it possible to perform object extraction with high analysis accuracy and little boundary distortion.

It is a block diagram which shows roughly the hardware constitutions of the image analysis apparatus of one Embodiment of this invention. It is a flowchart which shows roughly the process which an image analysis apparatus performs. It is a flowchart which shows an area | region label setting process roughly. It is a flowchart which shows roughly the clustering process using a subspace method. It is a flowchart which shows roughly the process of area | region boundary optimization using a likelihood function. It is explanatory drawing which shows an example of the attention texture extraction using a wavelet frame.

Explanation of symbols

1 Image analysis device 7 Storage medium

Claims (14)

Area dividing means for dividing the image into predetermined areas;
Multi-resolution analysis means for repeatedly executing multi-resolution analysis by a wavelet frame to a predetermined level for each of the areas divided by the area dividing means;
Feature extraction means for extracting a predetermined feature amount from the wavelet coefficients obtained by the multi-resolution analysis means;
Based on the feature amount extracted by the feature extraction unit, an error of the feature amount between neighboring regions is calculated, and the region and a region located in the vicinity of the region are determined according to the similarity of the calculated error By assigning the same label to a region label setting means for setting a label for each region,
A region with a label that matches a predetermined region is set as a selection region, a predetermined feature amount in the selection region is set as a cluster center, a partial space is defined in the selection region based on the cluster center, and the partial space and each Clustering means for measuring the degree of fitness with a region, and clustering into similar regions similar to the selected region according to the similarity of the fitness ,
A Markov random field model is introduced into each of the center region including the similar region composed of the selected region and the predetermined cluster obtained by the clustering, and the peripheral region that is a peripheral region thereof, and the likelihood of the region based on the model is determined. Region boundary optimization means for optimizing a boundary shape between the central region and the peripheral region by searching for a boundary having a maximum degree ;
An image analysis apparatus comprising: Recursively executing the region dividing unit, the multi-resolution analyzing unit, the feature extracting unit, and the region label setting unit until a predetermined end condition is satisfied,
The image analysis apparatus according to claim 1. Recursively executing the region dividing unit, the multi-resolution analyzing unit, the feature extracting unit, and the clustering unit until a predetermined end condition is satisfied,
The image analysis apparatus according to claim 1. The area dividing means divides a target image into a predetermined number of rectangular areas with a predetermined size,
The image analysis apparatus according to claim 1. The feature extraction means uses a vector having each of the average energy obtained for each predetermined component from each predetermined decomposition level as a feature amount.
The image analysis apparatus according to claim 1. Installed on the computer,
An area dividing function for dividing an image into predetermined areas;
A multi-resolution analysis function that repeatedly executes a multi-resolution analysis by a wavelet frame to a predetermined level for each of the areas divided by the area division function;
A feature extraction function for extracting a predetermined feature amount from a wavelet coefficient obtained by the multi-resolution analysis function;
Based on the feature amount extracted by the feature extraction function, an error of the feature amount between neighboring regions is calculated, and the region is located in the vicinity of the region according to the similarity of the calculated error A region label setting function for setting a label for each region by assigning the same label to
A region with a label that matches a predetermined region is set as a selection region, a predetermined feature amount in the selection region is set as a cluster center, a partial space is defined in the selection region based on the cluster center, and the partial space and each A clustering function that measures the degree of fitness with a region, and clusters the similar regions similar to the selected region according to the similarity of the fitness ,
A Markov random field model is introduced into each of the center region including the similar region composed of the selected region and the predetermined cluster obtained by the clustering, and the peripheral region that is a peripheral region thereof, and the likelihood of the region based on the model is determined. A region boundary optimization function that optimizes a boundary shape between the central region and the peripheral region by searching for a boundary having the maximum degree ;
Image analysis program that causes a computer to execute. Causing the computer to recursively execute the region division function, the multi-resolution analysis function, the feature extraction function, and the region label setting function until a predetermined end condition is satisfied,
The image analysis program according to claim 6 . Causing the computer to recursively execute the region division function, the multi-resolution analysis function, the feature extraction function, and the clustering function until a predetermined termination condition is satisfied,
The image analysis program according to claim 6 . Storage medium for storing the image analysis program according to any one of claims 6-8. An area dividing step for dividing the image into predetermined areas;
A multi-resolution analysis step for repeatedly executing a multi-resolution analysis by a wavelet frame to a predetermined level for each of the regions divided by the region division step;
A feature extraction step for extracting a predetermined feature amount from the wavelet coefficients obtained by the multi-resolution analysis step;
Based on the feature amount extracted by the feature extraction step, the error of the feature amount between neighboring regions is calculated, and the region is located in the vicinity of the region according to the similarity of the calculated error An area label setting step for setting a label for each area by giving the same label to
A region with a label that matches a predetermined region is set as a selection region, a predetermined feature amount in the selection region is set as a cluster center, a partial space is defined in the selection region based on the cluster center, and the partial space and each A clustering step of measuring the degree of fitness with a region and clustering into similar regions similar to the selected region according to the similarity of the fitness ,
A Markov random field model is introduced into each of the center region including the similar region composed of the selected region and the predetermined cluster obtained by the clustering, and the peripheral region that is a peripheral region thereof, and the likelihood of the region based on the model is determined. A region boundary optimization step for optimizing a boundary shape between the central region and the peripheral region by searching for a boundary having a maximum degree ;
An image analysis method comprising: Recursively executing the region dividing step, the multi-resolution analysis step, the feature extraction step, and the region label setting step until a predetermined end condition is satisfied,
The image analysis method according to claim 10 . Recursively executing the region dividing step, the multi-resolution analysis step, the feature extraction step, and the clustering step until a predetermined end condition is satisfied,
The image analysis method according to claim 10 . The region dividing step divides a target image into a predetermined number of rectangular regions with a predetermined size,
Image analysis method of any one of claims 10 to 12. The feature extraction step uses a vector having each of the average energy obtained for each predetermined component from each predetermined decomposition level as a feature amount.
Image analysis method of any one of claims 10 to 12.

Images