JP2015164012A - Category discriminator generation apparatus, category discrimination device, and computer program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To facilitate category discrimination even in a case of learning a similar category.SOLUTION: A category discriminator generation device comprises: a first image feature amount extraction unit extracting a first image feature amount from a learning image; a second image feature amount extraction unit extracting a second image feature amount; a similar category determination unit determining whether or not a learned category similar to a category of the learning image is present on the basis of the extracted second image feature amount and a second image feature amount of the learned category; a feature amount generation unit generating the second image feature amount of the similar, learned category and a dis-similar second image feature amount and generating a new second image feature amount in addition to the second image feature amount of the learning image on the basis of the first image feature amount of the learning image and a first image feature amount of the similar, learned category if it is determined that the learned category is present; and a category learning unit generating a category discriminator for discriminating the new second image feature amount.

Description

  The present invention relates to a technique for identifying a category of an image.

As a technique for identifying a category such as an object or a scene included in an image, for example, the following technique is available. Feature amounts are extracted from learning images with category labels in advance, and classifiers of a plurality of categories are learned using an SVM (Support Vector Machine) or the like. Then, the feature amount of the unknown image that is the category identification target is extracted, and it is determined whether the unknown image belongs to the learned category by using the learned classifiers of the plurality of categories. In this way, an unknown image category can be identified.
The feature quantities used for category identification can be roughly divided into two feature quantities. The first feature amount is a low-level feature amount (hereinafter referred to as “Low-level feature amount”) based on an image signal such as SIFT (Scale-Invariant Feature Transform) or a color histogram. The second feature amount is a mid-level feature amount newly learned from the low-level feature amount (hereinafter referred to as “Mid-level feature amount”). Mid-level feature amounts include, for example, feature amounts representing texture elements (for example, see Non-Patent Document 1) and urban areas such as tails, legs, furs in dogs, wings, windows, and metals in airplanes. This is a feature amount representing an object or environment (for example, see Non-Patent Document 2) constituting a scene such as a building, road, person, and a tree, grassland, or river in a natural landscape.

When low-level features are used for category identification, the categories are learned directly from the image signal, so the learning process is opaque and it is effective to use what features for what categories. It is difficult to verify whether it exists. In addition, since the semantic gap between the image signal and the category is large, a large amount of learning data is required for highly accurate identification.
When using mid-level features for category identification, first learn the intermediate element classifiers that make up the category based on the low-level feature quantities, and learn the categories using the output values of the intermediate element classifiers as feature quantities. Let Since the category is expressed by a combination of intermediate elements that can be perceived by humans, the learning process is clear when the mid-level feature is used for category identification. In addition, since the semantic gap between the Mid-level feature value and the category is smaller than the gap between the Low-level feature value and the category, high-precision identification is possible with a small amount of learning data.
However, when using the mid-level feature value, it is necessary to determine in advance which intermediate element should be used as the feature value. In Non-Patent Documents 1 and 2, an intermediate element is determined manually, and a classifier of the determined intermediate element is learned. In Non-Patent Document 3, when the category to be identified is known, the optimum feature value of the intermediate element is automatically acquired in order to discriminate the category to be identified.

A. Farhadi, I. Endres, D. Hoiem, and D. Forsyth. “Describing Objects by their Attributes”. In CVPR, pp. 1778-1785, 2009. L. Li, H. Su, E.P. Xing, and L. Fei-Fei. “Object Bank: A High-level Image Representation for Scene Classification & Semantic Feature Sparsification”. In NIPS, pp. 1378-1386, 2010. F.X. Yu, L Cao, R.S. Feris, J.R. Smith, and S.-F. Chang. “Designing Category-Level Attributes for Discriminative Visual Recognition”. In CVPR, pp. 771-778, 2013.

  However, when a category is newly learned, if the intermediate element of the known category matches the intermediate element of the newly learned category, the newly learned category cannot be distinguished. For example, let us consider a case where the Mid-level feature amount learned in advance is tail, foot, wing, the known category is dog / airplane, and a new category of cat is learned. When the cat category and the airplane category are compared, the cat category has a tail / foot but no wings, and the airplane category has wings but no tail / foot. Therefore, the category of the cat category and the airplane category can be discriminated because the intermediate elements are contradictory to each other. However, when comparing the cat category with the dog category, both categories have tails / feet and no wings. Therefore, in the cat category and the dog category, each intermediate element matches (collises), and it is difficult to discriminate between the two categories. As described above, when learning similar categories, there is a problem that it is difficult to discriminate between categories only by using Mid-level feature values learned in advance.

  In view of the above circumstances, an object of the present invention is to provide a technique capable of easily discriminating categories even when learning similar categories.

  According to one aspect of the present invention, a first image feature amount extraction unit that extracts a first image feature amount from a learning image for learning a category of an image, and a second from the extracted first image feature amount Similar to the category of the learning image based on the second image feature amount extraction unit that extracts the image feature amount, the extracted second image feature amount, and the second image feature amount of the learned category A similar category determining unit that determines whether or not the learned category exists, and the first image feature amount of the learned image when it is determined that the similar learned category exists. And generating a second image feature quantity that is not similar to the second image feature quantity of the similar learned category based on the first image feature quantity of the learned category to be generated, and The image feature amount of 2 is used as the learning image. In addition to the second image feature amount, a feature amount generation unit that generates a new second image feature amount, and a category learning unit that generates a category identifier for identifying the new second image feature amount And a category discriminator generating device.

  One aspect of the present invention is the above-described category classifier generation device, wherein the category learning unit is configured to use a relational expression of output values of a linear classifier for the learned category that is not similar to the category of the learning image. The weight for the second image feature amount newly added to the second image feature amount of the learning image is eliminated.

  One aspect of the present invention is a category identification device that identifies a category of an input image using the category identifier learned by the category learning unit of the category identifier generation device.

  One embodiment of the present invention is a computer program for causing a computer to function as the above-described device.

  According to the present invention, even when similar categories are learned, category discrimination can be facilitated.

It is a schematic block diagram showing the function structure of the category identifier production | generation apparatus 100 in this invention. It is a conceptual diagram regarding the collision detection of the Mid-level feature-value of a learning category. It is a conceptual diagram for demonstrating the process in case a similar category exists. It is a conceptual diagram for demonstrating the process in case a similar category exists. It is a schematic block diagram showing the function structure of the category identification apparatus 200 in this invention. It is a flowchart which shows the flow of a process of the category identification device production | generation apparatus 100 in this embodiment.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a schematic block diagram showing a functional configuration of a category discriminator generation device 100 according to the present invention. The category discriminator generation device 100 is a device that generates, by learning, a discriminator (category discriminator) that identifies an input image (category learning image) to be classified into what category. The category is a moving object or a scene included in the image, such as a cat or a dog. Further, the category learning image stored in the category learning image storage unit 10 is input to the category discriminator generation device 100. The category learning image is image data with a category indicating the image and the category of the image. The category learning image is stored in the category learning image storage unit 10 in association with an index for identifying the category learning image. The category discriminator generation device 100 learns a category discriminator based on the input category learning image.
Hereinafter, a specific configuration of the category discriminator generation device 100 will be described.

  The category discriminator generation device 100 includes a CPU (Central Processing Unit), a memory, an auxiliary storage device, and the like connected by a bus, and executes a category discriminator generation program. By executing the category discriminator generation program, the category discriminator generation device 100 has a low-level feature quantity extraction unit (first image feature quantity extraction unit) 101, a low-level feature quantity storage unit 102, a mid-level feature quantity extraction. Unit (second image feature amount extraction unit) 103, Mid-level feature amount storage unit 104, similar category determination unit 105, feature amount generation unit 106, category learning unit 107, and category classifier storage unit 108. . Note that all or part of each function of the category discriminator generating apparatus 100 is realized by using hardware such as an application specific integrated circuit (ASIC), a programmable logic device (PLD), and a field programmable gate array (FPGA). Also good. The category identifier generation program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a storage device such as a hard disk built in the computer system. The category identifier generation program may be transmitted / received via a telecommunication line.

The low-level feature quantity extraction unit 101 extracts a low-level feature quantity (first image feature quantity) from the category learning image of the input learning target category (hereinafter referred to as “learning category”). For example, SIFT, SURF (Speed-Up Robust Features), HoG (Histogram of oriented Gradient), color histogram, edge histogram, wavelet feature, and the like are applied to the extraction of the low-level feature amount.
The low-level feature amount storage unit 102 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The low-level feature quantity storage unit 102 stores the category learning image index of the learning category together with the low-level feature quantity extracted by the low-level feature quantity extraction unit 101.

The mid-level feature quantity extraction unit 103 extracts a mid-level feature quantity (second image feature quantity) from the low-level feature quantity extracted by the low-level feature quantity extraction unit 101. For the extraction of the mid-level feature quantity, the output value of any classifier other than the classifier of the learning category can be used. Examples of the mid-level feature amount include Attribute described in Non-Patent Document 1. More specifically, when the category is dog, airplane, or cat, the tail, paws, and wings correspond to Attribute.
The mid-level feature amount storage unit 104 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The mid-level feature amount storage unit 104 stores the index of the category learning image together with the mid-level feature amount extracted by the mid-level feature amount extraction unit 103.

  The similar category determination unit 105 determines whether there is a category similar to the learning category (hereinafter referred to as “similar category”) based on the Mid-level feature amount stored in the Mid-level feature amount storage unit 104. To determine. Specifically, the similar category determination unit 105 detects whether or not the Mid-level feature value of the learning category collides with the Mid-level feature value of the learned category. If a similar category exists, that is, if the learning category Mid-level feature amount and the learned category Mid-level feature amount collide, the similar category determination unit 105 characterizes the index information of the similar category. The data is output to the quantity generation unit 106. On the other hand, when the similar category does not exist, that is, when the Mid-level feature amount of the learning category does not collide with the Mid-level feature amount of the learned category, the similar category determination unit 105 determines the Mid-level of the learning category. The feature amount is output to the category learning unit 107. Specific processing of the similar category determination unit 105 will be described later.

  The feature amount generation unit 106 generates a new mid-level feature amount based on the index information of the similar category. The new Mid-level feature value is a Mid-level feature value that is not stored in the Mid-level feature value storage unit 104. Specifically, first, the feature quantity generation unit 106 stores the low-level feature quantity of the learning category and the low-level feature quantity of the similar category based on the index information of the similar category. 102. Next, the feature quantity generation unit 106 learns the classifier using the acquired low-level feature quantity of the learning category as a positive example and the low-level feature quantity of the similar category as a negative example. Thereby, the feature quantity generation unit 106 can discriminate between the learning category and the similar category. Then, the feature amount generation unit 106 generates the learned output value of the discriminator as a new Mid-level feature amount. For example, a linear SVM may be used as the discriminator. The new mid-level feature value generated by the feature value generation unit 106 is also used when learning a new category. That is, the feature value generation unit 106 newly stores the generated Mid-level feature value as a Mid-level feature value of each learned category stored in the Mid-level feature value storage unit 104.

  The category learning unit 107 learns a learning category classifier (category classifier). For example, when a Mid-level feature amount of a learning category is input from the similar category determination unit 105, the category learning unit 107 learns a learning category identifier based on the input Mid-level feature amount. For example, a linear classifier may be used for learning the classifier. Further, the category learning unit 107 learns the learning category discriminator by adding the new Mid-level feature amount obtained by the feature amount generating unit 106 to the Mid-level feature amount of the learning category. Here, the category classifier learned before generating a new Mid-level feature quantity and the category classifier learned after generating a new Mid-level feature quantity (learning category classifier) are Mid- The number of dimensions of level features is different. Therefore, re-learning is usually required for the former category classifier. However, the category learning unit 107 does not need to learn again by performing the following processing. The output value in the linear discriminator is determined by the inner product of the n-dimensional feature quantity a and the weight w of the feature quantity a as shown in Equation 1. Therefore, for learned categories that are not similar, the weight w_new of the feature quantity generated by the feature quantity generation unit 106 in Formula 2 becomes the same formula as Formula 1, so that the identification result remains unchanged. Can do. Therefore, re-learning of the discriminator becomes unnecessary.

  The category identifier storage unit 108 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The category classifier storage unit 108 stores the classifier (category classifier) learned by the category learning unit 107.

FIG. 2 is a conceptual diagram related to collision detection of the Mid-level feature amount of the learning category.
FIG. 2 shows, in order from the left, the category learning image, the Mid-level feature value, and the collision detection result between the learning category Mid-level feature value and the learned category Mid-level feature value. In FIG. 2, images related to the categories “airplane”, “dog”, and “cat” are shown as category learning images. The categories “airplane” and “dog” correspond to learned categories, and the category “cat” Corresponds to the learning category. In the example of FIG. 2, three values of “Wing”, “Tail”, and “Furry” are shown as the Mid-level feature amount. The category learning images 1-1 to 1-3, 2-1 to 2-3, and 3-1 to 3-3 for each category “airplane”, “dog”, and “cat” have respective Mid-level feature amounts. The category learning images 1-1 to 1-3, 2-1 to 2-3, and 3-1 to 3-3 are shown in association with each other. For example, the mid-level feature amount of the category learning image 1-1 of the category “airplane” is “A ¹ ₁ = {0.90, 0.21, 0.18}”. The similar category determination unit 105 determines the learning category (FIG. 2) based on the mid-level feature amounts of the respective category learning images 1-1 to 1-3, 2-1 to 2-3, and 3-1 to 3-3. Then, it is determined whether there is a similar category of the category “cat”.
Hereinafter, specific processing of the similar category determination unit 105 will be described.

The similar category determination unit 105 performs a two-sample test on the Mid-level feature quantity of the learning category and the Mid-level feature quantity of the learned category in order to detect a collision, and whether or not the null hypothesis is rejected. Test. A t-test may be used for the statistical test. The similar category determination unit 105 performs a two-sample test on each dimension of the mid-level feature quantity (in FIG. 2, “Wing”, “Tail”, “Furry”), and when the null hypothesis is rejected Determines that the dimensions for which the null hypothesis is rejected do not collide. On the other hand, when the null hypothesis is not rejected, the similar category determination unit 105 determines that the dimensions for which the null hypothesis is not rejected are colliding.
Then, the similar category determination unit 105 calculates the collision rate R based on the determination result for each dimension. Specifically, the similar category determination unit 105 calculates the collision rate R using the collision rate R = N_collion / N, where N is the dimension of the Mid-level feature value and N_collion is the number of collided dimensions.

  When the collision rate R exceeds the threshold, the similar category determination unit 105 determines that the mid-level feature quantities of both categories cannot be discriminated, that is, there is a similar category. In this case, the similar category determination unit 105 outputs the index information of the similar category to the feature amount generation unit 106. On the other hand, when the collision rate R does not exceed the threshold value, the similar category determination unit 105 determines that the mid-level feature quantities of both categories can be discriminated, that is, there is no similar category. The threshold value may be stored in advance in the category discriminator generation device 100 or may be arbitrarily determined by the user.

In the example shown in FIG. 2, the collision detection result between the category “airplane” and the category “cat” has a collision rate of 0%. That is, it is shown that the category “airplane” and the category “cat” can be distinguished. In this case, the similar category determination unit 105 outputs the mid-level feature amount of the category learning image of the category “cat” to the category learning unit 107. In the example shown in FIG. 2, the collision detection result of the category “dog” and the category “cat” has a collision rate of 100%. That is, it is shown that the category “dog” and the category “cat” cannot be distinguished (collision). In this case, the similar category determination unit 105 outputs the index information of the similar category (category “dog”) to the feature amount generation unit 106.
Above, the description about the specific process of the similar category determination part 105 is complete | finished.

  Next, specific examples of operations of the similar category determination unit 105 and the feature amount generation unit 106 will be described with reference to FIGS. 3 and 4 are conceptual diagrams for explaining the processing in the case where a similar category exists. In the description of FIGS. 3 and 4, a case where the classifiers of the categories “dog” and “airplane” have been learned in advance will be described as an example. In addition, for the sake of simplification of explanation, an attribute (for example, “Wing” or “Tail”) in which a moving body (for example, “dog” or “airplane”) indicated by a category is indicated by a mid-level feature amount is used. If it has, it is indicated by “1”, otherwise it is indicated by “0”.

  In the example shown in FIG. 3A, the item corresponding to the attribute “Wing” of the learned category “dog” is “0”, and the item corresponding to the attribute “Tail” is “1”. That is, it is shown that the dog does not have wings (“Wing”) but has a tail (“Tail”). In addition, the item corresponding to the attribute “Wing” of the learned category “airplane” is “1”, and the item corresponding to the attribute “Tail” is “0”. That is, it is shown that the airplane has a wing and does not have a tail.

  As described above, the value of each attribute is different between the learned categories “dog” and “airplane”. Therefore, the similar category determination unit 105 determines that the collision rate R is less than the threshold for the learned categories “dog” and “airplane”. That is, the Mid-level feature value of the learned category “dog” and the Mid-level feature value of “Airplane” do not collide. In this case, since the learned categories “dog” and “airplane” are mutually contradictory, when the classifier of the category “cat” is newly learned, there is only one set of learned categories that collide.

When a new classifier “category” is trained, a mid-level feature value (“Wing” or “Tail”) based on the low-level feature value extracted from the category learning image corresponding to the category “cat”. ) Is extracted. FIG. 3B shows an example in which items corresponding to the attributes of the learning category “cat” are added.
In the example shown in FIG. 3B, the item corresponding to the attribute “Wing” of the learning category “cat” is “0”, and the item corresponding to the attribute “Tail” is “1”. That is, the cat does not have wings (“Wing”) but has a tail (“Tail”).

  As shown in FIG. 3B, the value of each attribute is the same in the learned category “dog” and the learning category “cat”. Therefore, the similar category determination unit 105 determines that the collision rate R is greater than or equal to the threshold value for the learned category “dog” and the learned category “cat”. That is, the Mid-level feature value of the learned category “dog” and the Mid-level feature value of the learning category “Cat” collide. In this case, the feature value generation unit 106 newly adds a feature value of an attribute (New) that can distinguish both categories (learned category “dog” and learning category “cat”) as a Mid-level feature value (see FIG. 3 (C)). Since a feature amount of an attribute that can discriminate between the learned category “dog” and the learned category “cat” is newly added, all categories are contradicted again. Then, a discriminator for identifying the learning category “cat” is generated.

Further, when the classifier of the category “Lion” is newly learned again, since all categories are contradictory, there is only one set of conflicting categories. When a classifier of category “Lion” is newly learned, a Mid-level feature amount (“Wing” or “Tail”) based on the low-level feature amount extracted from the category learning image corresponding to category “Lion”. Or “New”) is extracted. FIG. 4A shows an example in which items corresponding to the attributes of the learning category “lion” are added.
In the example shown in FIG. 4A, the item corresponding to the attribute “Wing” of the learning category “Lion” corresponds to “0”, the item corresponding to the attribute “Tail” corresponds to “1”, and the attribute “New”. The item is “1”. That is, the lion does not have a wing (“Wing”), has a tail (“Tail”), and has an attribute “New”.

As shown in FIG. 4A, the value of each attribute is the same in the learned category “Cat” and the learning category “Lion”. Therefore, the similar category determination unit 105 determines that the collision rate R is greater than or equal to the threshold value for the learned category “cat” and the learned category “lion”. That is, the Mid-level feature value of the learned category “Cat” and the Mid-level feature value of the learning category “Lion” collide. In this case, the feature quantity generation unit 106 newly adds a feature quantity of an attribute (New2) that can distinguish both categories (learned category “cat” and learning category “lion”) as a Mid-level feature quantity (FIG. 4 (B)). Since a feature amount of an attribute that can discriminate between the learned category “cat” and the learned category “lion” is newly added, all categories are contradicted again. Then, a discriminator for identifying the learning category “lion” is generated. Each time a new category is learned, the above processing is repeatedly executed.
Above, the description about the specific operation example of the similar category determination part 105 and the feature-value production | generation part 106 is complete | finished.

Hereinafter, the operation in the category discriminator generation device 100 will be described using a specific example.
The category discriminator generating apparatus 100 learns categories one by one (sequentially), that is, sequentially generates category discriminators for identifying a certain category. As a specific example, an example in which a discriminator for discriminating dogs, airplanes, and cats is sequentially learned will be described. Further, as the mid-level feature value, for example, the feature value of the tail, the foot, or the wing is used.
First, the category discriminator generation device 100 learns a discriminator of the learning category “dog”. When a category learning image in which a dog is imaged is input to the category discriminator generation device 100, the low-level feature value extraction unit 101 extracts a low-level feature value from the category learning image. Next, the Mid-level feature quantity extraction unit 103 extracts Mid-level feature quantities (tail, foot, wing) from the Low-level feature quantities.

  The similar category determination unit 105 determines whether a similar category exists based on the Mid-level feature amount stored in the Mid-level feature amount storage unit 104. At this time, since there is no similar category, the category learning unit 107 learns a discriminator of the category “dog” based on the mid-level feature amount (tail, foot, wing) of the learning category “dog”.

  Next, the category discriminator generation device 100 learns the discriminator of the learning category “airplane”. When a category learning image in which an airplane is imaged is input to the category discriminator generation device 100, the low-level feature quantity extraction unit 101 extracts a low-level feature quantity from the category learning image. Next, the Mid-level feature quantity extraction unit 103 extracts a Mid-level feature quantity from the Low-level feature quantity.

  The similar category determination unit 105 determines whether a similar category exists based on the Mid-level feature amount stored in the Mid-level feature amount storage unit 104. In this case, the similar category determination unit 105 determines whether or not the mid-level feature value of the learning category “airplane” and the mid-level feature value of the learned category “dog” collide with each other. Since the mid-level feature amount of the learning category “airplane” and the mid-level feature amount of the learned category “dog” are contradictory, there is no collision. That is, since there is no similar category, the category learning unit 107 learns a classifier of “airplane” based on the mid-level feature amount of the learning category “airplane”.

  Next, the category discriminator generation device 100 learns the discriminator of the learning category “cat”. When a category learning image in which a cat is captured is input to the category discriminator generation device 100, the low-level feature amount extraction unit 101 extracts a low-level feature amount from the category learning image. Next, the Mid-level feature quantity extraction unit 103 extracts a Mid-level feature quantity from the Low-level feature quantity.

  The similar category determination unit 105 determines whether a similar category exists based on the Mid-level feature amount stored in the Mid-level feature amount storage unit 104. In this case, the similar category determination unit 105 determines whether or not the Mid-level feature amount of the learning category “Cat” and the Mid-level feature amount of the learned category “Airplane” collide with each other. Further, the similar category determination unit 105 determines whether or not the mid-level feature amount of the learning category “cat” and the mid-level feature amount of the learned category “dog” collide with each other.

  Since the mid-level feature amount of the learning category “cat” and the mid-level feature amount of the learned category “airplane” are contradictory, there is no collision. On the other hand, the Mid-level feature amount of the learning category “Cat” and the Mid-level feature amount of the learned category “Dog” collide because the collision rate R is equal to or greater than the threshold value. In this case, the feature value generation unit 106 generates a new mid-level feature value (New) based on the low-level feature value of the learning category “cat” and the low-level feature value of the learned category “dog”. To do. Specifically, the linear discriminator is learned with the low-level feature value of the learning category “cat” as a positive example and the low-level feature value of the learned category “dog” as a negative example. The category learning unit 107 first learns the discriminator of the learning category “cat” based on the four Mid-level feature values of “tail”, “foot”, “wing”, and “New”. Further, the category learning unit 107 re-learns the discriminator of the learned category “dog”, which is a similar category, based on the four mid-level feature quantities, similarly to the discriminator of the learning category “cat”.

  After that, when learning a new category, the category discriminator generating apparatus 100 sets four Mid-level feature amounts newly added by the above-described processing to four Mid-level feature amounts as targets of the new category. Extracted from the captured category learning image. Then, the category discriminator generation device 100 determines whether or not the extracted Mid-level feature amount collides with the Mid-level feature amount of each of the learned categories “dog”, “airplane”, and “cat”. If a similar category exists, a new Mid-level feature (New2) is added.

FIG. 5 is a schematic block diagram showing the functional configuration of the category identification device 200 according to the present invention. The category identification device 200 identifies the category of the image input to the category identification device 200 using the category identifier generated by the category identifier generation device 100. For example, the category identification apparatus 200 receives a test image stored in the test image storage unit 20. Unlike the category learning image, the test image does not include information other than the image. The category identification device 200 identifies the category of the input test image using the category identifier generated by the category identifier generation device 100.
Hereinafter, a specific configuration of the category identification device 200 will be described.

  The category identification device 200 includes a CPU, a memory, an auxiliary storage device, and the like connected by a bus, and executes a category identification program. By executing the category identification program, the category identification device 200 functions as a device including a low-level feature quantity extraction unit 201, a mid-level feature quantity extraction unit 202, and a category classifier 203. Note that all or some of the functions of the category identification device 200 may be realized using hardware such as an ASIC, PLD, or FPGA. The category identification program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a storage device such as a hard disk built in the computer system. The category identification program may be transmitted / received via a telecommunication line.

The low-level feature quantity extraction unit 201 extracts a low-level feature quantity from the test image.
The mid-level feature quantity extraction unit 202 extracts a mid-level feature quantity from the low-level feature quantity extracted by the low-level feature quantity extraction unit 201.
The category identifier 203 is a category identifier learned by the processing of the category learning unit 107. The category discriminator 203 identifies the category of the input test image based on the Mid-level feature amount extracted by the Mid-level feature amount extraction unit 202.

FIG. 6 is a flowchart showing a process flow of the category discriminator generation device 100 in the present embodiment.
A category learning image is input to the category discriminator generation device 100 (step S101). When the category learning image is input, the low-level feature amount extraction unit 101 extracts the low-level feature amount from the input category learning image (step S102). Thereafter, the low-level feature quantity extraction unit 101 stores the extracted low-level feature quantity in the low-level feature quantity storage unit 102 together with the index of the category learning image.

  Next, the mid-level feature quantity extraction unit 103 extracts a mid-level feature quantity from the low-level feature quantity extracted by the low-level feature quantity extraction unit 101 (step S103). Thereafter, the Mid-level feature quantity extraction unit 103 stores the extracted Mid-level feature quantity in the Mid-level feature quantity storage unit 104 together with the index of the category learning image.

The similar category determination unit 105 performs similar categories based on the Mid-level feature amount extracted by the Mid-level feature amount extraction unit 103 and the Mid-level feature amount stored in the Mid-level feature amount storage unit 104. Is determined (step S104). When the similar category does not exist (step S104—NO), the category learning unit 107 learns the learning category classifier based on the learning category mid-level feature amount (step S105). Thereafter, the category learning unit 107 stores the learned classifier in the category classifier storage unit 108 (step S106).
On the other hand, when a similar category exists (step S104—YES), the feature quantity generation unit 106 generates a new mid-level feature quantity (step S107).

  According to the category discriminator generation device 100 configured as described above, a highly accurate category discriminator can be generated. Specifically, a learned category similar to the learning category of the category learning image input to the category discriminator generation device 100 is automatically determined, and the feature amount of the attribute that can distinguish the similar category from the learning category Are sequentially generated and added as Mid-level features. Therefore, there is at most one Mid-level feature amount added when a category is newly learned. As described above, the category discriminator generation device 100 can classify categories by the minimum number of attributes without increasing a large number of attributes necessary for category identification by learning the categories sequentially. it can. Therefore, even when learning similar categories, category discrimination can be facilitated.

  In addition, as in Non-Patent Document 3, in the batch method of acquiring the optimum feature for all categories, it is necessary to renew the feature every time a category is added, and according to the increase in the number of categories to be added. As a result, the calculation amount increases. On the other hand, since the present invention generates only the feature quantity that can discriminate between the learning category and the similar category, the calculation amount is less likely to increase with the increase in the number of categories. Therefore, it is efficient to learn a category discriminator.

In addition, when the category discriminator generating apparatus 100 learns a discriminator of a learning category by adding a new Mid-level feature quantity, the number of dimensions of the learned category discriminator and the Mid-level feature quantity is different. Re-learning becomes necessary. However, as in the present invention, for learned categories that are not similar, re-learning can be made unnecessary by calculating the weight for the newly added Mid-level feature amount as zero.
In addition, according to the category identification device 200 configured as described above, an image to be identified is identified using the category identifier generated by the category identifier generation device 100. Therefore, it becomes possible to identify the category of the image with high accuracy.

<Modification>
The category learning image input to the category discriminator generation device 100 may be one image or a plurality of images. Further, the test image input to the category identification device 200 may be one sheet or a plurality of sheets.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes designs and the like that do not depart from the gist of the present invention.

DESCRIPTION OF SYMBOLS 10 ... Category learning image memory | storage part, 20 ... Test image memory | storage part, 100 ... Category discriminator production | generation apparatus, 101 ... Low-level feature-value extraction part, 102 ... Low-level feature-value memory | storage part, 103 ... Mid-level feature Quantity extraction unit, 104 ... Mid-level feature quantity storage part, 105 ... Similar category determination part, 106 ... Feature quantity generation part, 107 ... Category learning part, 108 ... Category discriminator storage part, 200 ... Category discrimination device, 201 ... Low-level feature quantity extraction unit, 202 ... Mid-level feature quantity extraction unit, 203 ... Category identifier

Claims (4)

A first image feature amount extraction unit that extracts a first image feature amount from a learning image for learning an image category;
A second image feature quantity extraction unit for extracting a second image feature quantity from the extracted first image feature quantity;
Similarity for determining whether the learned category similar to the category of the learning image exists based on the extracted second image feature amount and the second image feature amount of the learned category A category determination unit;
When it is determined that there is a similar learned category, the learning is similar based on the first image feature quantity of the learning image and the first image feature quantity of the similar learned category. A second image feature amount that is not similar to the second image feature amount of the learned category is generated, and the generated second image feature amount is added to the second image feature amount of the learning image to newly A feature amount generation unit for generating a second image feature amount;
A category learning unit for generating a category discriminator for identifying the new second image feature amount;
A category discriminator generating device.   The category learning unit adds the second newly added to the second image feature amount of the learning image in the relational expression of the output value of the linear classifier for the learned category that is not similar to the category of the learning image. The category discriminator generation device according to claim 1, wherein a weight for the image feature amount is eliminated.   A category identification device that identifies a category of an input image using the category identifier learned by the category learning unit of the category identifier generation device according to claim 1.   The computer program for functioning a computer as an apparatus of any one of Claims 1-3.

Images