JP2014106565A - Object area boundary estimation device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the accuracy in extraction of an area and a boundary of an object included in an image.SOLUTION: An object area boundary estimation device includes: a small area division unit which divides the whole of each of images included in an image data set into small areas; a local area feature extraction unit which extracts a local area feature vector from each small area; a global area feature extraction unit which extracts a global area feature vector from each small area; a local boundary feature extraction unit which extracts a local boundary feature vector for each of boundaries between pairs of adjacent small areas; a global boundary feature extraction unit which extracts a global boundary feature vector for each of boundaries between pairs of adjacent small areas; a quantization unit which outputs a quantization feature vector through discretization about local area feature vectors, global area feature vectors, local boundary feature vectors, and global boundary feature vectors; a model estimation unit which learns an area/boundary multiplex feature model with the quantization feature vector as an observation variable: and a topic estimation unit which estimates the area and boundary of an object in the image on the basis of the learning result.

Description

  The present invention relates to a technique for extracting an object region from an image and simultaneously estimating the relationship between the regions, and extracting an object region from a large amount of image data with high accuracy and classifying boundary lines between the regions. And the program.

  There are mainly two methods for estimating an object region from an image. The first method is a method of classifying the inside of an image with objects included in the image in advance, constructing a discriminator by learning, and discriminating using the discriminator (for example, Non-Patent Document 1). The second method is a method of constructing a probability model and learning probabilistically by learning a large amount of images labeled with object regions (for example, Non-Patent Document 2).

  Specifically, in the first method, a feature quantity is extracted from an image, and a classifier that identifies each class using the feature quantity as an input is configured. Examples of commonly used feature values include Bag-of-Features. In this case, a code book is created by clustering feature amounts extracted from a plurality of images, and the image is expressed with the appearance frequency.

  The first method is effective when each image is labeled and each image can be represented by one noun. If the variance of feature vectors obtained from images with the same label is small, the classifier can be configured with high accuracy, which is effective.However, if the image contains multiple objects, the same object may have several features in the feature space. It is difficult to apply when forming a cluster.

  On the other hand, in the second method, a feature quantity and a label generation probability model are constructed by learning a large amount of labeled image sets instead of learning feature quantities for each class of object regions. Image indexing is performed by estimating model parameters for a new input image.

  The second method is effective when an image includes a plurality of objects, or when several clusters are formed in the feature space even with the same object, and the region division and the region labeling are performed simultaneously. Can do.

Image classification based on Bag-of-features using co-occurrence representation of foreground and background information, Fujiyoshi Nagahara, “Image Recognition and Understanding Symposium (MIRU2010)” Current status and future of general object recognition, Keiji Yanai, IPSJ Journal CVIM19-No. SIG 16, Vol. 48,2007

  However, in the prior art using the second method, each image is treated as a set of local features, each as a local feature that has no relationship, or each image is treated as a set of small regions, Since it was treated as a collection of small areas, it was impossible to relate image areas. That is, based on the adjacency relationship between objects, it is not possible to perform region division and region labeling at the same time.

  As described above, in the related art, when a plurality of objects are included in an image, it is difficult to determine the similarity of images based on the adjacent relationship between objects. Here, the adjacency relationship between objects refers to the property of the boundary between the object and the object on the image related to the spatial positional relationship between the objects (regions are adjacent in the image but not in real space) , A boundary on the convex side of the orthogonal surface, a boundary on the concave side of the orthogonal surface, etc.).

  On the other hand, when the first method according to the prior art is applied, there is a problem that learning cannot be performed when there are a plurality of objects in an image and the adjacent relationship changes. In addition, when the second method according to the related art is applied, there is a problem that although it is possible to estimate a model including a plurality of objects, it is not possible to estimate the adjacent relationship between objects.

  The present invention has been made in view of such circumstances, and an object of the present invention is to provide an object region boundary estimation apparatus and program capable of improving the accuracy of extracting an object region and an object boundary in an image. And

  In order to solve the above problem, the object region boundary estimation apparatus of the present invention includes a first small region dividing unit that divides the entire image into small regions for each image included in the image data set, and a first small region dividing unit for each small region. A first local region feature extraction unit that extracts one local region feature vector; a first global region feature extraction unit that extracts a first global region feature vector for each of the small regions; and the two adjacent small regions A first local boundary feature extraction unit that extracts a first local boundary feature vector for each boundary line between regions, and a first global boundary feature vector for each boundary line between two adjacent small regions Discretizing the first global boundary feature extraction unit, the first local region feature vector, the first global region feature vector, the first local boundary feature vector, and the first global boundary feature vector First quantum A first quantization unit that outputs a feature vector, a first model estimation unit that learns a region boundary multiple feature model using the first quantization feature vector as an observation variable, and a first model estimation unit And a topic estimation unit that estimates an object region and a boundary in an image of input image data based on a learning result.

  Further, the present invention is the object region boundary estimation apparatus described above, wherein the topic estimation unit includes a second small region dividing unit that divides the entire image of the input image data into small regions, and the first A second local region feature extraction unit that extracts a second local region feature vector for each small region divided by the two small region dividing units; and for each small region divided by the second small region dividing unit A second global region feature extraction unit that extracts a second global region feature vector; and a second for each boundary line between two adjacent small regions among the small regions divided by the second small region dividing unit. A second local boundary feature extracting unit for extracting a local boundary feature vector of the second sub-region, and a second global region for each boundary line between two adjacent small regions among the small regions divided by the second small region dividing unit Second global boundary feature that extracts boundary feature vectors An extraction unit, the second local region feature vector, the second global region feature vector, the second local boundary feature vector, the second global boundary feature vector, and the first model estimation unit. And a second model estimation unit that classifies each small region divided by the second small region dividing unit and a boundary line between each small region based on a learning result.

  In order to solve the above problem, the present invention is a program for causing a computer to function as the object region boundary estimation apparatus described above.

  According to the present invention, learning is performed on a small region obtained by dividing an input image and a boundary between the small regions based on a model using two types of a global feature and a local feature. The accuracy of extracting the object region and the object boundary in the image can be improved.

It is a schematic block diagram which shows the structure of the object area | region boundary estimation apparatus 1 in this embodiment. It is a figure which shows the relationship of each feature vector which the feature vector extraction part 120 extracts in the embodiment. It is a figure which shows the concept of the area boundary multiple feature model which the model estimation part 13 uses in the embodiment. It is a graphical model which shows the stochastic relationship between each element (each parameter and each topic). It is a figure which shows the application example which estimated the area | region topic and the boundary topic with respect to the image of several horses by the object area | region boundary estimation apparatus 1 of the embodiment.

  Hereinafter, an object region boundary estimation apparatus 1 according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a schematic block diagram showing a configuration of an object region boundary estimation apparatus 1 in the present embodiment. The object region boundary estimation apparatus 1 includes a parameter learning unit 10 and a topic estimation unit 20. The parameter learning unit 10 receives an image data set including a plurality of images to be learned. The parameter learning unit 10 outputs various model parameters as learning results. Image data indicating one image is input to the topic estimation unit 20. The topic estimation unit 20 outputs a topic associated with the small region and the boundary line in the input image.

  The parameter learning unit 10 includes a small region dividing unit 11, a visual word generating unit 12, and a model estimating unit 13. The visual word generation unit 12 includes a feature vector extraction unit 120, a quantization dictionary storage unit 125, and a quantization unit 126. The feature vector extraction unit 120 includes a local region feature extraction unit 121, a global region feature extraction unit 122, a local boundary feature extraction unit 123, and a global boundary feature extraction unit 124. The model estimation unit 13 includes a region boundary multiple feature model storage unit 131.

  The topic estimation unit 20 includes a small region division unit 21, a visual word generation unit 22, and a model estimation unit 23. The visual word generation unit 22 has the same configuration as the visual word generation unit 12 in the parameter learning unit 10. The model estimation unit 23 includes a region boundary multiple feature model storage unit 231 and a model parameter storage unit 232.

  The operation of the object region boundary estimation apparatus 1 shown in FIG. 1 will be described with reference to FIG. In the parameter learning unit 10, the small region dividing unit 11 divides the input image indicated by the image data into a large number of small regions for each image data of an image data set including a plurality of input image data. This small region assumes that the same object is captured in the small region. The small area dividing unit 11 divides the input image sufficiently finely so that a plurality of objects are not included in the same small area. As a method for dividing an input image into small regions, for example, “Graph-based Image Segmentation” by Felzenszwalb et al., “Entropy Rate Superpixel” by Liu et al., And the like can be used.

  In the visual word generation unit 12, the feature vector extraction unit 120 performs local region feature, global region feature, local boundary feature, and global boundary feature for each small region of the image divided by the small region dividing unit 11. Generate 4 types of visual words. The local region feature extraction unit 121, the global region feature extraction unit 122, the local boundary feature extraction unit 123, and the global boundary feature extraction unit 124 generate a feature vector as described below as a visual word. The relationship between each feature vector is shown in FIG.

  FIG. 2 is a diagram showing the relationship between the feature vectors extracted by the feature vector extraction unit 120 in the embodiment. In the figure, as an example, an image including three apples placed on a plate and one apple placed outside the plate is divided into small areas. Has been. Moreover, the example which extracts each characteristic with respect to the apple put out of the plate among four apples is shown.

  The local area feature extraction unit 121 extracts a local area feature vector indicating a local area feature from an input image divided into small areas. The local region feature indicates a local region feature vector extracted from feature points randomly obtained from the divided small regions. Specifically, the local region feature extraction unit 121 extracts an image feature amount expressing a local gradient such as a SIFT (Scale Invariant Feature Transform) feature amount or a SURF (Speeded Up Robust Features) feature amount. The local region feature extraction unit 121 selects feature points to be extracted randomly at a constant rate with respect to the region area.

  The global area feature extraction unit 122 extracts a global area feature vector indicating a global area feature from the entire area divided into small areas. The global area feature is a feature vector indicating the pixel distribution of the entire divided small area. The global area feature extraction unit 122 extracts a global area feature vector using, for example, a color histogram, barycentric coordinates, texture features, and the like.

  The local boundary feature extraction unit 123 extracts a local boundary feature vector indicating the local boundary feature for the boundary line between the divided small regions. The local region feature is extracted from points randomly selected on the boundary line with respect to the boundary line between adjacent small regions among the divided small regions. The local boundary feature extraction unit 123 uses, for example, “GeometricBlur” as the local region feature on the boundary line (on the edge). The local boundary feature extraction unit 123 selects feature points from which local boundary features are extracted at random with respect to the length of the boundary line.

  The global boundary feature extraction unit 124 extracts a global boundary feature vector indicating the global boundary feature for the boundary line between the divided small regions. The global boundary feature is a feature vector indicating a gradient distribution of one whole line connecting two adjacent small regions with respect to a boundary line between adjacent small regions among the divided small regions. The global boundary feature extraction unit 124 extracts a global boundary feature vector using, for example, a gradient histogram such as an EOH (Edge of Orientation Histograms) feature. Note that the global boundary feature extraction unit 124 can similarly extract the global boundary feature vector using any feature related to the entire line such as the length and inclination of the boundary line.

  The quantization dictionary storage unit 125 includes a quantization dictionary in which local region features, global region features, local boundary features, and clusters to which each feature vector of the global boundary features belongs and an index indicating the cluster are associated with each other. Remembered. The cluster to which each feature vector belongs is constructed by clustering the four types of feature vectors.

  The quantization unit 126 quantizes each of the four feature vectors extracted by the feature vector extraction unit 120 using the quantization dictionary stored in the quantization dictionary storage unit 125. The quantization unit 126 outputs an index corresponding to the cluster to which each feature vector belongs as a quantized visual word. As a clustering technique, for example, K-means clustering (K-average method) can be used. The quantization unit 126 outputs the quantized vector, which is each quantized feature vector, to the model estimation unit 13 as a visual word.

  The model estimation unit 13 uses the visual word output from the visual word generation unit 12 to learn a parameter indicating the stochastic relationship between each small region and boundary and each visual word. In addition, the model estimation unit 13 outputs a parameter that characterizes the object region and the boundary model in the image as a learning result. The concept of the region boundary multiple feature model used for learning by the model estimation unit 13 is shown in FIG.

FIG. 3 is a diagram showing a concept of a region boundary multiple feature model used by the model estimation unit 13 in the same embodiment. A plurality of small areas are obtained from the input image, and a boundary is set between adjacent small areas. From each small region, one global region feature (x ^(g) ) and a plurality of local region features (x ₁ ^(l) , x ₂ ^(l) , x ₃ ^(l) ,...) ^Are extracted. From each boundary between the small regions, one global boundary feature (y ^(g) ) and a plurality of local boundary features (y ₁ ^(l) , y ₂ ^(l) , y ₃ ^(l) ,...) ^Are extracted. The The region boundary multiple feature model storage unit 131 stores a parameter indicating each extracted feature and a parameter indicating a stochastic relationship described later.

  FIG. 4 is a graphical model showing a probabilistic relationship between each element (each parameter and each topic). Each small region boundary is defined for all adjacent regions. However, in the graphical model shown in the figure, the relationship between a pair of adjacent small regions is extracted and described for simplicity.

In the figure, the area topic of each small area is indicated by T, and the boundary topic of each boundary is indicated by U. Observed visual words are shown as gray circles, and parameters given for learning are shown as black dots. The visual words output by the visual word generation unit 12 are represented by the global region feature x ^(g) , the local region feature x ^(l) , the global boundary feature y ^(g) , and the local boundary feature y ^(l) , respectively. Has been.

As parameters indicating the probabilistic relationship between topics and features, there are φ ^(g) , φ ^(l) , η ^(g) , and η ^(l) . φ ^(g) indicates the relationship between global region features and region topics, φ ^(l) indicates the relationship between local region features and region topics, and η ^(g) indicates the relationship between global boundary features and boundary topics. , Η ^(l) indicates the relationship between local boundary features and boundary topics. Further, a vector indicating the appearance frequency of the area topic in the entire image is indicated by θ, and the relationship between the adjacent area topic and the boundary topic is indicated by ξ.

~ Φ ^(g) , ~ φ ^(l) , ~ η ^(g) , ~ η ^(l) , ~ θ, and ~ ξ are φ ^(g) , φ ^(l) , η ^(g) , η ^{(l )} , Parameters of the Dirichlet distribution (prior distribution) for generating θ, ξ. In FIG. 4, a character with “˜ (tilde)” added above θ is described as “˜θ” in the specification. The same applies to other characters (φ, η, etc.).

In FIG. 4, N _I , K, and L described at the lower right of the frame indicate the number of images, the number of area topics, and the number of boundary topics, respectively. N _I , K, and L mean that there are a plurality of the insides of the frame. There are also a plurality of relationships between region boundaries surrounded by broken lines in the image.

Such a probability model is called a generation model. Hereinafter, the generation process in the probability model will be described. First, Dirichlet distributions θ, φ, η, and ξ having parameters ~ θ, ~ φ, ~ η, and ~ ξ are generated. θ represents the probability that each small region belongs to each label. A region label T is generated according to the probability θ. Next, the parameters φ ^(g) and φ ^(l) indicate the probability that a visual word of each small area is generated according to each area label. Small region visual words x ^(g) , x ^(l) are generated according to the parameters φ ^(g) , φ ^(l) and the region labels.

Next, the parameter ξ is the appearance probability of the boundary label according to the combination of adjacent region labels. A boundary label is generated by the combination of the parameter ξ and the adjacent region label. Finally, boundary visual words y ^(g) and y ^(l) are generated by parameters η ^(g) and η ^(l) indicating the appearance probability of the boundary feature corresponding to the boundary label.

  A posteriori distribution of each variable is obtained according to this model, and parameters φ, η, and ξ are learned. “Variational Message Passing” proposed by John Winn et al. Is used as a method for obtaining the posterior distribution of each variable. “Variational Message Passing” is a technique for updating the posterior distribution of each node by propagating the expected value to the adjacent node for each node of the graphical model in the framework of inference called variational Bayesian method. “Variational Message Passing” can be applied universally to a probability model described by a probability distribution belonging to an exponential distribution genus.

According to “Variational Message Passing”, the update formula of each variable in the model of the parameter learning unit 10 is as follows. The update of θ _k indicating the frequency of appearance of the topic topic k follows the following equation (1).

In Equation (1), <•> indicates an expected value of a variable, and ψ (•) indicates a digamma function. t _rk indicates the probability that the r-th region T _r is a topic k, and N _r indicates the number of small regions in the image. The area topic T is expressed by the probability t _rk that the small area r becomes the topic k, and is updated by the following equation (2).

In Expression (2), β (r) indicates an index of a local region feature in the region r, and ε (r) indicates an index of a region adjacent to the region r. (R, s) indicates the boundary between the region r and the region s. Also, ξ _kk′l indicates the probability that the boundary between the topic k and topic k ′ regions will be topic l.

Similarly, the boundary topic U is expressed by the probability u _bl that the boundary b becomes the topic l and is updated by the following equation (3).

Here, the boundary b is a boundary between the region r ₁ and the region r _2, and β (b) indicates an index of a local boundary feature in the boundary b. Since φ ^(g) , φ ^(l) , η ^(g) , and η ^(l) are updated by the same formula from the relationship between each visual word and each topic, the update formula for φ ^(g) is shown as φ ^{( g)} Description of the update formulas of ^l) , η ^(g) , and η ^(l) is omitted. The following equation (4) is an update equation for φ ^(g) .

  In equation (4), δ (a, b) represents a Dirac δ function that is 1 only when a = b, and V represents the maximum value that the visual word can take.

ξ is updated by the following equation (5). Here, (r, s) represents the boundary between the region r and the region s, N _b indicates the number of boundaries in the image.

  Update is performed a sufficient number of times for each node, and the obtained posterior distributions of φ, η, and ξ are used as model parameters in the topic estimation unit 20. That is, the parameters φ, η, and ξ calculated by the parameter learning unit 10 are stored in the model parameter storage unit 232 included in the topic estimation unit 20.

  Next, the operation in the topic estimation unit 20 will be described. The topic estimation unit 20 is input with target image data to be calculated. Similar to the small area dividing unit 11 in the parameter learning unit 10, the small area dividing unit 21 divides the image indicated by the input image data into a large number of small areas. Similar to the visual word generation unit 12 in the parameter learning unit 10, the visual word generation unit 22 applies a local region feature, a global region feature, a local boundary feature, and a global region to the image divided by the small region division unit 21. Four types of visual words with boundary features are generated. Since the visual word generation unit 22 has the same configuration as the visual word generation unit 12, description of the configuration of the visual word generation unit 22 is omitted.

  The model estimation unit 23 outputs a topic for each small region and boundary using the visual word extracted by the visual word generation unit 22. The region boundary multiple feature model storage unit 231 stores a visual word indicating each feature extracted by the visual word generation unit 22 as a parameter. The model parameter storage unit 232 stores the posterior distributions of φ, η, and ξ output from the parameter learning unit 10 as ~ φ, ~ η, and ~ ξ indicating the prior distribution. That is, the model estimation unit 23 calculates the region topic T and the boundary topic U for the input image based on the learning result (the posterior distribution of φ, η, and ξ) by the parameter learning unit 10.

Similarly to the parameter learning unit 10, the model estimation unit 23 applies each variable (φ ^(g) , φ ^(l) , η ^(g) , η ^(l) ) by applying “Variational Message Passing”. Then, the topic topic T and the boundary topic U are calculated by maximizing these values. The formula used when updating each variable is the same as the formula used in the description of the model estimation unit 13 in the parameter learning unit 10. The model estimation unit 23 repeats updating a sufficient number of times for each node, and outputs the value having the highest posterior probability as the estimation topic.

  As described above, in the parameter learning unit 10, the visual word generation unit 12 performs the local region feature, the global region feature, the local boundary feature, and the global boundary feature for the small region divided by the small region dividing unit 11. The four types of visual words are generated. A visual word is obtained by quantizing an image feature into several discrete values by clustering. The expression using the visual word makes it possible to express an image by a probability model using a discrete distribution. The visual word generation unit 12 generates two types of visual words (global region feature and local region feature, and global boundary feature and local boundary feature) for each region and boundary. The model estimation unit 13 can learn the same object and the same boundary line as the same topic in the relationship between the visual word and the topic by giving the same topic to a plurality of visual words obtained from the same region and boundary. it can.

  As described above, the model estimation unit 13 divides the boundary feature into the global boundary feature and the local boundary feature, extracts a plurality of local boundary features from one boundary, and learns them as the same boundary topic. The model estimation unit 13 can accurately classify boundary lines having similar meanings such as the contour of an object, an internal edge, and a ridge line by learning using such a region boundary multiple feature model. In other words, the model estimation unit 13 learns the region boundary multiple feature model using the visual word (quantized vector obtained by quantizing each feature vector) output from the visual word generation unit 12 as an observation variable.

  Further, by using parameters (a posteriori distribution of φ, η, and ξ) that are parameters that are output as learning results by the model estimation unit 13 having the characteristics as described above and that characterize the object region and boundary models, topics can be obtained. The estimation unit 20 can classify regions and boundaries in the input image with high accuracy.

  FIG. 5 is a diagram illustrating an application example in which the object topic boundary estimation apparatus 1 according to the embodiment estimates a region topic and a boundary topic for a plurality of horse images. This is an area topic and boundary topic estimated by the topic estimation unit 20 based on parameters obtained by inputting 300 horse images as an image data set to the parameter learning unit 10 and based on parameters obtained by the parameter learning unit 10. The same area topic is shown with the same shading pattern, and the same boundary topic is shown with the same shading pattern. As shown in the figure, it can be seen that two types of horse regions and horse contours are extracted as the same topic.

  1 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read and executed by a computer system, whereby the learning described above is performed. Alternatively, the area topic and the boundary topic may be calculated. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer system” includes a WWW system having a homepage providing environment (or display environment). The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Further, the “computer-readable recording medium” refers to a volatile memory (RAM) in a computer system that becomes a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. In addition, those holding programs for a certain period of time are also included.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

  As mentioned above, although one Embodiment of this invention was described with reference to drawings, said Embodiment is only the illustration of this invention, and it is clear that this invention is not limited to the said Embodiment. Accordingly, additions, omissions, substitutions, and other modifications of the components may be made without departing from the technical idea and scope of the present invention.

  For example, the small region dividing unit 11 and the small region dividing unit 21 in FIG. 1 have the same configuration, and the visual word generating unit 12 and the visual word generating unit 22 have the same configuration. Therefore, the object region boundary estimation apparatus 1 includes a small region dividing unit 11, a visual word generation unit 12, a model estimation unit 13, and a model estimation unit 23, and the visual word generation unit 12 performs model estimation when learning. The visual word may be output to the unit 13, and the visual word generation unit 12 may output the visual word to the model estimation unit 23 when calculating the area topic and the boundary topic.

  With regard to technology that extracts object regions from images and simultaneously estimates the relationship between regions, it can also be applied to applications where image processing that extracts object regions from large amounts of image data with high accuracy and classifies boundary lines between regions is essential. .

  DESCRIPTION OF SYMBOLS 1 ... Object area | region boundary estimation apparatus, 10 ... Parameter learning part, 11 ... Small area division | segmentation part, 12 ... Visual word generation part, 120 ... Feature vector extraction part, 121 ... Local area | region feature extraction part, 122 ... Global area | region feature extraction part , 123 ... local boundary feature extraction unit, 124 ... global boundary feature extraction unit, 125 ... quantization dictionary storage unit, 126 ... quantization unit, 13 ... model estimation unit, 131 ... region boundary multiple feature model storage unit, 20 ... topic Estimating unit, 21 ... small region dividing unit, 22 ... visual word generating unit, 23 ... model estimating unit, 231 ... region boundary multiple feature model storage unit, 232 ... model parameter storage unit

Claims (3)

A first small region dividing unit that divides the entire image into small regions for each image included in the image data set;
A first local region feature extraction unit that extracts a first local region feature vector for each small region;
A first global region feature extraction unit that extracts a first global region feature vector for each of the small regions;
A first local boundary feature extraction unit that extracts a first local boundary feature vector for each boundary line between two adjacent small regions;
A first global boundary feature extraction unit that extracts a first global boundary feature vector for each boundary line between two adjacent small regions;
Discretizing the first local region feature vector, the first global region feature vector, the first local boundary feature vector, and the first global boundary feature vector, and outputting a first quantized feature vector A first quantization unit;
A first model estimator that learns a region boundary multiple feature model using the first quantized feature vector as an observation variable;
A topic estimation unit that estimates an area and a boundary of an object in an image of input image data based on a learning result of the first model estimation unit;
An object region boundary estimation apparatus comprising: The object region boundary estimation apparatus according to claim 1,
The topic estimation unit includes:
A second small area dividing unit for dividing the entire image of the input image data into small areas;
A second local region feature extracting unit that extracts a second local region feature vector for each small region divided by the second small region dividing unit;
A second global area feature extraction unit that extracts a second global area feature vector for each of the small areas divided by the second small area dividing unit;
A second local boundary feature extraction unit that extracts a second local boundary feature vector for each boundary line between two adjacent small regions among the small regions divided by the second small region dividing unit;
A second global boundary feature extraction unit that extracts a second global boundary feature vector for each boundary line between two adjacent small regions among the small regions divided by the second small region dividing unit;
Based on the second local region feature vector, the second global region feature vector, the second local boundary feature vector, the second global boundary feature vector, and a learning result of the first model estimation unit A second model estimation unit that classifies each small region divided by the second small region dividing unit and a boundary line between each small region;
An object region boundary estimation apparatus comprising:   The program for functioning a computer as an object area | region boundary estimation apparatus as described in any one of Claim 1 or Claim 2.