JP6448212B2 - Recognition device and recognition method - Google Patents

Description

  The present invention relates to a recognition apparatus and a recognition method, and in particular, image information such as a still image, a moving image, or a distance image is used as an input image, and image attribute information such as a scene, event, composition, or main subject is used as a clue. The present invention relates to a technique suitable for use in estimation.

  As a conventional method for estimating an image scene or event using an object in an image as a clue, for example, there is Non-Patent Document 1. Non-Patent Document 1 examines whether or not a plurality of specific classes of objects are present in an image, and performs image scene discrimination using the distribution of the result of the presence or absence of the presence or absence as a feature amount.

Li-Jia Li, Hao Su, Longwan Lim, Li Fei-Fei, “Objects as Attributes for Scene Classification”, Proc. of the European Conf. on Computer Vision (ECCV 2010). P. Felzenszwalb, R.A. Girstick, D.M. McAllester, and D.M. Ramanan. "Object Detection with Discriminative Trained Part Based Models", IEEE Trans. on Pattern Analysis and Machine Intelligence 2010 Koen E.M. A. van de Sande, Jasper R .; R. Uijlings, Theo Gevers, Arnold W. M.M. Smeulders, Segmentation As Selective Search for Object Recognition, IEEE International Conferencing on Computer Vision, 2011 Jianbo Shi and Jitendra Malik, Normalized Cuts and Image Segmentation, IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. 22, no. 8, 2000 Joao Carreira and Christian Sminchiscu, Constrained Parametric Min-Cuts for Automatic Object Segmentation, IEEE Conference on Computer Revision and Pattern Recognition 20 Svetlana Lazebnik, Cordelia Schmid, Jean Ponce, Beyond Bags of Features 6 P. Felzenszwalb, D.W. McAllester, D.M. Ramanan, A Discriminative Trained, Multiscale, Deformable Part Model, IEEE Conferencing on Computer Vision and Pattern Recognition, 2008 T.A. Kobayashi and N.K. Otsu. Action and Simultaneous Multiple-Person Identification Using Cubic Higher-Order Local Auto-Correlation ", In Proc. International Conference on Patter.

  In such a method, a plurality of detectors (for example, using a technique such as Non-Patent Document 2) for recognizing a subject as a clue for scene discrimination are prepared, and detection processing is performed for each specific subject such as a dog or a car. There is a need. At this time, the following problems exist.

  First, in order to accurately determine the types of a large number of scenes, it is necessary to prepare detectors for a large number of subjects related to each scene. The processing of the detector as in Non-Patent Document 2 requires a large amount of calculation because image scanning called a sliding window is performed for each detector, and the processing time required for scene discrimination significantly increases as the number of scenes increases. there is a possibility.

  Second, since it is generally unknown which subject is important when determining a scene, it is difficult to determine in advance what kind of subject detector should be prepared when identifying a subtle scene. is there.

Third, for example, when differentiating between the scenes of a birthday party and a wedding reception, a difference in variation, not the presence or absence of a subject, may be important, such as a difference in whether a person's clothes are usually worn or dressed. In the conventional method as in Non-Patent Document 1, there is a problem that the difference in the variation of the subject cannot be used as a clue for discrimination as it is.
In view of the above-described problems, an object of the present invention is to make it possible to determine a scene of an image by a method with a low processing load using various subjects as clues.

Recognition device of the present invention, from the input image, the candidate area extraction means for extracting a plurality of candidate areas of the object, a feature amount extracting section which extracts a feature quantity of the candidate region from each of said plurality of candidate areas, the for each of the plurality of candidate regions, on the basis of the feature amount extracted from the candidate regions, with reference to the learning result that associates the feature quantity extracted from the object region of the learning image and the attribute of the learning image , Attribute determination means for determining the attribute of the input image from which the candidate area is extracted, and identification means for identifying the attribute of the input image by aggregating the determination results of the attribute determination means for the plurality of candidate areas It is characterized by having.

  According to the present invention, it is possible to discriminate an image scene by a method with a low processing load using various subjects as clues. Further, it is not necessary to teach in advance what kind of subject should be identified in each scene.

It is a block diagram which shows the basic composition of the recognition apparatus of 1st Embodiment. It is a flowchart explaining the flow of a process of the recognition apparatus of 1st Embodiment. It is a flowchart explaining the procedure of the process which extracts a candidate area | region. It is a flowchart explaining the procedure of the process which extracts the characteristic of a candidate area | region. It is a schematic diagram of an attribute determination processing unit. It is a flowchart explaining the procedure of the determination process of an attribute. It is a figure which shows the example of a result of the determination process of an attribute. It is a block diagram which shows the basic composition of the learning phase of 1st Embodiment. It is a flowchart explaining the procedure of the process of the learning phase of 1st Embodiment. It is a figure which shows the example of a result of extraction of the object area for learning in a learning phase. It is a flowchart explaining the procedure of the learning process of the classification tree in a learning phase. It is a schematic diagram of the learning result of an attribute determination process part. It is a block diagram which shows the structural example of the derivative form of 1st Embodiment. It is a block diagram which shows the basic composition of the recognition apparatus of 2nd Embodiment. It is a block diagram which shows the basic composition of the recognition apparatus of 3rd Embodiment. It is a figure which shows the example of extraction of the candidate area | region of the object of 3rd Embodiment. It is a block diagram which shows the basic composition of the recognition apparatus of 4th Embodiment. It is a figure which shows the example of a composition class and an estimation result. It is a figure which shows the example of the estimation result of the main subject area.

[First Embodiment]
Hereinafter, an example of a mode for carrying out the recognition apparatus of the present invention will be described with reference to the drawings. An object of the present recognition apparatus is to receive an input image and correctly determine to which of a plurality of predetermined scene classes the image belongs.

  FIG. 1 shows a basic configuration diagram of a recognition apparatus according to the present embodiment. The image input unit 101 inputs image data. The candidate area extraction unit 102 extracts a subject area related to the attribute of the image for determining the attribute information of the image. The feature amount extraction unit 103 extracts an image feature amount from the candidate region of the subject extracted by the candidate region extraction unit 102. The attribute determination unit 104 determines, based on the feature amount extracted by the feature amount extraction unit 103, a candidate region that is included in an image of any scene class. The determination result integration unit 105 collects the results of the attribute determination unit 104 and finally determines the scene class of the image.

  Here, the scene class of the image includes various scenes and event types such as a birthday party, a Christmas party, a wedding, a camp, an athletic meet, and a school performance. In this embodiment, it is assumed that several tens of classes of scenes as described above are given by the user, and these scene classes are discriminated from the input image.

  Furthermore, the present invention can be applied to a class other than a scene class related to a daily event such as the above-described example. For example, shooting set for the purpose of adjusting shooting parameters when the camera performs shooting, such as night view / light source direction (forward light, backlight, oblique light from the right, oblique light from the left), flower close-up, etc. A type called a mode can also be defined and used as a scene class. As described above, the present invention is applicable to discrimination of attributes of various target images.

Next, the flow of recognition processing and learning processing of the recognition apparatus of the present embodiment will be described.
<Recognition phase>
First, the flow of recognition processing will be described using the flowchart of FIG.
In S201, the image input unit 101 receives image data. Here, the image data in the embodiment of the present invention refers to various types of video information such as a color image, a moving image, or a distance image, and combinations thereof. In the present embodiment, it is assumed that a color image of a still image is input to the image input unit 101. Further, in this step, preprocessing necessary for the subsequent recognition processing such as size enlargement / reduction and luminance value normalization is performed as necessary according to the image.

  Next, in S202, the candidate area extraction unit 102 performs a process of extracting candidate areas of a plurality of subjects that are clues for identifying the scene class of the image. The subject mainly includes a region of an object having a certain shape and size such as a human body, a dog, and a cup, and a region related to a background having a relatively large size and an indefinite shape such as the sky, lawn, and a mountain.

  In the present embodiment, a mode in which image scenes are classified by analyzing objects among them will be described. For example, if there is an object area that seems to be a dish in the image, the possibility of a party scene is high, and if there is an object shaped like a lamp, it is highly likely that it is a camping scene. Condition.

  It is desirable for the candidate area extraction unit 102 to extract the entire area of the object existing in the image as accurately as possible. This is because, when a scene is determined based on an object, if the extracted region includes only a part of the object or different objects are mixed, the possibility of erroneous determination increases.

  However, it is extremely difficult to accurately cut out the boundaries of individual objects without knowing what the objects are about various objects present in the image. Therefore, here, a plurality of candidate regions that are considered to be object regions are extracted without expecting perfect object region segmentation. Then, assuming that some of them include an object region with a certain degree of accuracy, the image scene is determined in each candidate region. It is hoped that the majority of the results will finally determine the scene class correctly even if some errors are included.

  There are various known methods that can be used to extract candidate regions under such conditions, but in this embodiment, the method of Non-Patent Document 3, which is a known technique, is used as a reference. The flow of processing will be briefly described with reference to the flowchart of FIG.

First, in S301, the image is divided into small regions in which pixels having a similar color called Super-pixel (hereinafter referred to as SP) are grouped.
Next, in S302, (1) texture feature similarity and (2) size similarity are calculated for all adjacent SP pairs. A value obtained by weighting and summing this with a predetermined coefficient α (0 ≦ α ≦ 1) as in Equation 1 is used as the SP pair similarity.

  Here, as a texture feature, a frequency histogram of a general color SIFT feature is used as a feature amount (refer to Non-Patent Document 3 for details). As the similarity of the texture features, histogram intersection, which is a widely used distance measure, is used. The SP size similarity is a value obtained by dividing the area of the smaller SP of the SP pairs by the area of the larger SP. Both the texture similarity and the size similarity are 0 when the similarity is minimum and 1 when the similarity is maximum.

  Next, in S303, the pair of two adjacent SPs having the highest similarity is connected to form a subject candidate area. The connected SP is set as a new SP, and the similarity between the feature amount and the adjacent SP is recalculated. This is repeated until all the SPs are connected (S302 to S306). In this way, as many and small candidate areas as the number obtained by subtracting 1 from the number of SPs are extracted. However, since a subject area that is too small is likely to cause erroneous determination, only SPs having an area larger than a predetermined value are set as candidate areas (S304 to S305). Refer to Non-Patent Document 3 for details of the above candidate region generation processing.

  In addition, although it demonstrated with reference to the method of nonpatent literature 3 here, the method of extracting the area | region like an object is not restricted to this, Various methods can be considered. For example, a technique for separating the foreground and background called graph cut, or a technique for performing texture analysis to divide an image into textures (refer to Non-Patent Document 4 for details) is widely known.

Returning to the flowchart of FIG.
In S203, the feature amount extraction unit 103 extracts a feature amount from each candidate area. The feature quantity to be extracted includes a plurality of types. Here, six types of feature amounts are extracted as shown in the flowchart of FIG.

(1) In S401, a frequency histogram of color SIFT features in the candidate area is extracted.
(2) In S402, the area and position of the candidate region are output. Here, the value of the area is a value obtained by normalizing the whole image as 1, and the position is a barycentric position of a region normalized by setting the vertical and horizontal lengths of the image to 1.

  (3) In S403, a color SIFT feature is extracted from the area around the candidate area to generate a frequency histogram. This is because the feature of the background is also a clue when recognizing the object, so that the feature of the surrounding region is given separately. The area around the candidate area is an area expanded by a predetermined width of the candidate area.

Next, the following fourth to sixth feature amounts are calculated in steps S404 to S406 as feature amounts that serve as clues for determining whether the candidate region is a true object region.
(4) In S404, the similarity between the frequency SIFT features in the candidate region and the frequency histogram of the color SIFT features around the candidate region is calculated as a feature amount. Here, histogram intersection is used as the similarity.

(5) In S405, the average value of the edge strength of the candidate region contour is calculated as the feature amount. Here, the absolute value (dx ² + dy ² ) ^1/2 of the luminance gradient of the image is calculated on the contour of the region to obtain the average value. However, dx and dy are values of luminance gradients in the x and y directions of the image, respectively.

(6) In S406, the convex shape degree of the candidate area is calculated as a feature amount. Here, the convex shape degree is a ratio between the area of the candidate region and the area of the convex hull of the candidate region. This value is 1 if the candidate area is convex, and approaches 0 if the candidate area is concave.
By using the characteristics as described above, it is possible to determine to some extent whether or not the candidate area is an object area independent from the surrounding area.

  The above is the feature extraction process performed in S203. In addition to the feature values used here, various feature values related to the region such as the aspect ratio and moment length of the candidate region can be used. It should be noted that the embodiment of the present invention is not limited to the above-described feature amount.

  Next, in S204, the attribute determination unit 104 determines which scene class is associated with the feature of each candidate region. Note that it is an important feature of the present invention that what is determined here is not the type of the candidate area itself but the type of the scene class to which the candidate area relates.

  In the present embodiment, the attribute determination unit 104 includes an ensemble classification tree 502 configured by a set (ensemble) of a plurality of classification trees 502a to 502c, as schematically illustrated in FIG. Each classification tree estimates the correct scene class from the feature quantity of the candidate area, and votes the estimation result to the scene class voting space of FIG.

  Each discrimination node (partially indicated by symbols 503a to 503c) includes a linear discriminator. Each leaf node (partially indicated by symbols 503d to 503f) stores information on learning data assigned to the leaf node in the learning phase. Symbols 504a to 504c are attached to a part of the learning data, and symbols 505a to 505c are attached to the labels of the corresponding scene classes. As will be described in detail later in the learning phase, each discriminating node is learned so that candidate areas of learning data are divided into scene classes as much as possible.

The procedure of the determination process using the ensemble classification tree will be described with reference to the flowchart of FIG.
When the candidate area is input to the ensemble classification tree in S601, the determination process is performed for each classification tree in S603 to S607.

First, in S603, the scene class discrimination process is started from the root node 503a of the classification tree. In each node, discrimination by a discriminator is performed based on the feature of the candidate area.
In S604, the next branch is determined based on the determination result, and the process proceeds.
In S605, it is determined whether or not the leaf node has been reached. If not, the processing returns to S604, and the determination and movement in each node are repeated until the leaf node is reached. When the leaf node is reached, the scene class ratio of the learning data stored in the leaf node is referred to. This is the likelihood score of the scene class in the input candidate area.

However, if the node reaches a leaf node with a high ratio of learning data (which is non-object region data, which is called non-object region class) with the symbol φ, the input candidate region is not an object region. Probability is high. Therefore, voting is not performed from a leaf node having non-object region class data exceeding a predetermined ratio (S606). Otherwise, the likelihood score value is voted and added to each class. (S607). In the example of FIG. 5A, since the ratio of the birthday party class is 2/3 in the reached leaf node, a value of 0.667 is added.
There are several other variations of the voting method. For example, it is possible to select one scene class with the maximum ratio of leaf nodes and vote only one vote.

As described above, when all the classification trees have determined and voted for all candidate regions, the voting process S205 is ended.
In step S206, the scene class having the maximum total number of votes thus obtained is output as the scene class of the input image. However, as another form, when the number of votes of any scene class does not fall within a predetermined threshold, it may be output that the scene class is unknown. Conversely, if the number of votes for a plurality of scene classes exceeds a predetermined threshold value at the same time, all the corresponding scene classes may be output.

  FIG. 7 shows an example of results from extraction of candidate areas to scene class determination. FIG. 7A shows an example of an input image. FIG. 7B shows a candidate area of an object extracted by the candidate area extraction unit 102 with a rectangular frame (however, all candidate areas are not shown in order to avoid complexity).

  FIG. 7C is a diagram after removing candidate areas determined to be non-object area classes as a result of having all candidate areas determined by the attribute determination unit 104. FIG. 7D shows an example of the result of voting a scene class from the remaining candidate areas in this way. The symbol θ in the figure is a threshold value for determining the scene class. In this example, since the party scene exceeds the threshold θ, the party scene is output and the recognition process is terminated.

<Learning phase>
Next, processing for learning an ensemble classification tree, called a learning phase, will be described.
The purpose of this process is to provide the attribute determination unit 104 with the following three items: (1) a learning image set provided by the user, and (2) a teacher value of a scene class and (3) a teacher value of an object position. To give and train. Then, an ensemble classification tree that can correctly output the scene class type for the input image is created.

  FIG. 8 shows a configuration diagram of the recognition apparatus in the learning phase. This is similar to FIG. 1 of the basic configuration diagram in the recognition phase. The difference from FIG. 1 is that an object position data input unit 106 and an image attribute data input unit 107 exist and input a teacher value of an image scene class and an object position.

The learning phase operation process will be described with reference to the flowchart of FIG.
First, in S901, the image input unit 101 inputs a learning image. At the same time, the image attribute data input unit 107 inputs the teacher value of the scene class corresponding to each learning image.

In step S <b> 902, the object position data input unit 106 inputs a teacher value of the position of the object region for each learning image.
The object position data here is the one prepared by the user as the teacher value for the position of the object region included in the image for each learning image, as shown in FIG. In FIG. 10A, a rectangle circumscribing the object area is shown as an example of the form of the teacher value of the position of the object area (the symbols 1002a to 1002d are attached to a part of the rectangle of the teacher value. ).

  Note that it is difficult for the user to determine which object greatly contributes to the determination of the scene when the number of determination target scenes is large. For this reason, the user does not perform extraneous estimation for each object, but only determines whether the object is an object, and teaches as many object positions as possible. However, teaching may be omitted for objects that are too small, or objects that are occluded and not visible, to make learning difficult.

  Next, in S903, the candidate area extraction unit 102 extracts candidate areas by the same method as that performed in the recognition phase. FIG. 10B shows an example of the result. Here, symbols 1003a to 1003c are attached to some of the rectangles of the candidate areas.

  Next, in S904, the degree of overlap (overlap value) of the closest teacher-valued object region is examined for the extracted candidate region. The overlap value of the region x and the region y is calculated by the following formula 2.

Next, in S905, a candidate area whose overlap value with the object area is a predetermined value or more (here, 0.5 or more) is adopted as the object area used for learning. The object regions adopted for learning are shown with symbols 1004a to 1004e in FIG.
Further, the candidate area whose overlap value is less than a predetermined value (here, less than 0.2) is used as a non-object class area for subsequent learning. However, since the number of areas of the non-object class is larger than the number of areas of the object, the number is appropriately sampled and used.

  In step S <b> 906, the attribute determination unit 104 performs learning using the object region learning data and the non-object region learning data obtained by the above processing. Hereinafter, since it is an important process, a particularly detailed flow will be described with reference to FIG.

First, in S1101 of FIG. 11, the feature amount extraction unit 103 extracts feature amounts from all object regions and non-object regions by the same method as in the recognition phase. Hereinafter, in S1102 to S1106, feature amounts are extracted for each region.
In S1103, learning is started from the root node. First, a scene class in which learning data exists is randomly divided into two groups. Further, the non-object region is also regarded as an independent scene class, and either one of the two groups is selected at random and allocated together.

  In step S1104, the classifier is trained so that it can be discriminated into the two groups defined above using the feature amount of each learning data as input. Here, a general linear support vector machine (hereinafter referred to as SVM) method is used as a machine learning method. At this time, a part of the learning data is randomly sampled and is not used as the data for evaluation but used for learning.

  In the next S1105, the effectiveness of the discrimination capability of the SVM is evaluated using this evaluation data. First, the evaluation data is discriminated by SVM and divided into two. Then, the amount of information of the following (formula 3) that is generally used for learning the classification tree is calculated.

Here, c is a scene class variable, and p _kc is the k-side data (in this case, the divided data is indicated by L and R symbols) of the two divided data. It means the proportion of class c. N means the number of data before division, and _nk means the number of data divided on the k side. When the data is divided into two by the discriminator, the value of the information amount is large if there is a deviation in the class distribution in the data of each group after the division, and the value of the information amount is small if there is no deviation.
In S1102 to S1106, the definition of two random groups, the learning of SVM, and the evaluation of the discrimination result of SVM are repeated a predetermined number of times.

When the predetermined number of times is over, in S1107, the SVM parameter that has obtained the largest amount of information obtained so far is adopted as the discriminator of this node.
In S1108, the learning data is determined again using the adopted SVM learning parameters, and is divided into two depending on whether the SVM score is positive or negative, and each is assigned to the nodes on the right and left branches. Then, the processing from S1102 is recursively repeated on each of the left and right branches.

  However, as a result of the division, when there is only one kind of scene class (S1109), or when the number of data after division is less than a predetermined number (S1110), learning of the branch is terminated and a leaf node is obtained. . And the information of the learning data remaining at that time is memorize | stored in a leaf node, and a recursive process is complete | finished.

  In this manner, scene classes are classified by performing discrimination that causes deviation in appearance frequency of scene classes for each division at each node. A schematic example of such a learning result is shown in FIG.

  In FIG. 12, each leaf node of the classification tree 502a (partially indicated by the symbols 503d to 503f) is close to the feature quantity of the form, or data in the same area of the scene class is gathered. Although a complete classification is not necessarily performed, a high-accuracy scene class determination result can be obtained by learning a large number of such classification trees and integrating them by voting as an ensemble classification tree.

  Also note here that birthday cake 504b and Christmas cake 504c are on separate leaf nodes 503e and 503d. In this method, the subject is not classified by type such as cake or hat, but is classified by the scene class of the image to which the subject belongs. Therefore, even if an object of the same type as a cake belongs to different scene classes of birthday and Christmas and has different appearance variations, classification into different branches is automatically performed as shown in FIG. It is also shown that it can be performed. This is particularly emphasized here as it is an important effect of the embodiment of the present invention.

  Here, the SVM is used for learning of the discriminator, and the information amount criterion is used as the evaluation criterion. However, the learning method and the evaluation criterion are not particularly limited in the present invention, and a known discriminator technique is used. Widely applicable. For example, linear discriminant analysis may be used instead of SVM, and a general Gini coefficient may be used as an evaluation criterion.

  The above is the method of learning the classification tree. Only the learning of a single classification tree has been described here. The ensemble classification tree is composed of a plurality of classification trees, and when learning other classification trees, it is necessary to provide variations for each classification tree. There are a plurality of known methods as the method, but the most general method is used here. That is, learning data may be subsampled for each classification tree, and learning may be performed based on different learning data sets for each tree.

<Derived Form for Teaching Method of Subject Position>
In the above description, the learning method for giving the subject position as a teacher value has been described. However, teaching of the subject position at the time of learning is not an essential requirement in the present invention. In order to show that the scope of application of the present invention is not limited to this form, another derivative form of the subject teaching method will be described below.

The following four methods can be considered as examples of the method of teaching value of a subject.
(1) The user teaches a large number of object positions for all images.
(2) The user teaches only a part of the position of the subject.
(3) The user does not teach the position of the subject.
(4) Teach only the position of the subject that the user thinks is particularly relevant to the scene class.

  Here, (1) is the method already described. (4) indicates only objects that are significantly related to the scene, such as a Christmas tree at a Christmas party. Hereinafter, the derivation forms (2) to (4) will be briefly described in order.

First, an example (2) of teaching only a part of the position of the subject will be described. Here, only the differences from the embodiment (1) will be described.
The form example (2) is similar to a known learning method called semi-teacher learning. Specifically, first, feature quantities of the object region are extracted from the learning image in which the object position is taught. Next, candidate areas are extracted from learning images to which no teacher value is given, and their feature values are obtained. Next, if there is a region whose feature quantity is similar to a certain value or more from any of the taught regions, the candidate region is preferentially adopted as an object region for learning. The area below the predetermined value is used as learning data for the non-object class. The above is the description of the method of the embodiment (2).

  Next, an example (3) in which no subject information is taught will be described. In this embodiment (3), the classification tree is learned using all candidate regions extracted from the learning image without discrimination between object region data and non-object regions. The flow of the learning phase is as shown in the flowchart of FIG.

  In the flowchart of FIG. 9B, S911 corresponds to S901, S912 corresponds to S903, and S913 corresponds to S906. In this embodiment, the previously used “non-object region class” is not set. Therefore, incompletely cut out regions and regions unrelated to the object are also used for learning as subject regions related to the scene class.

In the method of the embodiment (3), although the burden on the user's teaching is reduced, the discrimination accuracy is low accordingly. There are a number of ideas for countermeasures.
First, as a first device, the number of data and the number of classification trees are increased, and the number of votes at the time of recognition is increased. As a general property of ensemble learning, it is widely known that even if the discrimination accuracy of each weak classifier (classification tree) is low, the discrimination accuracy increases asymptotically as the number of weak classifiers increases if the variation is large. .

  Furthermore, as another additional contrivance of the form (3) of the subject teaching method, the extent to which the object region is correctly extracted after the candidate region extraction unit 102 extracts the candidate region (hereinafter referred to as this). Is called object degree). An area determined to have a low object degree is not used for learning or recognition. Various methods for estimating the object degree are known, and for example, the method described in Non-Patent Document 5 can be cited.

  Non-Patent Document 5 solves the problem of regression that estimates the overlap value, which is the amount of overlap between the candidate area and the true object area, using the feature quantity of the candidate area as input (see Non-Patent Document 5 for details). )

Based on this, the processing flow is modified as shown in the flowchart of FIG.
That is, after extracting a candidate area in S922, an overlap value is estimated to be an object degree in S923, and an area having an object degree lower than a predetermined value is removed in S924, and then learning and recognition are performed in S925.

  Next, a method of the embodiment (4) for teaching only an important subject will be described. One example of implementation of the method (4) is as follows. In other words, the object candidate region that overlaps the important object specified by the user is weighted and learned by weighting the amount corresponding to the overlap amount when learning the classification tree. Then, other object candidate areas that do not overlap are learned without weighting as usual.

  Hereinafter, the method of the embodiment (4) will be described in detail. First, all object candidate regions are extracted as in the method of the embodiment (3) in which the object position is not taught. Furthermore, an object degree is estimated and the area | region with a low object degree is remove | excluded. Next, for all remaining regions, the overlap value with the important object designated by the user is calculated. Next, based on the overlap value, the object candidate region x is weighted with w (x) of the following Equation 4 to learn the classification tree.

Here, O (x) is the degree of overlap between the region x and the important object, and β is a coefficient equal to or greater than 0. If this value is large, the importance object is emphasized more strongly and learning is performed. For β, an appropriate value is set by the intersection confirmation method or the like.
Next, the information amount of Equation 4 is expanded, and Equation 3 is expanded and defined as follows as an information amount criterion considering the importance weight of learning data. What is necessary is just to learn a discriminator using this.

However, p _kc (w) is a weighted ratio of the number of learning data of scene class c,

It is. Thereby, it is possible to learn so as to distinguish the area close to the important object with more importance. Note that the above equation has the same value as the previous learning equation when β is 0 or when an important object does not exist in the learning case.
As above, the derivation examples (1) to (4) of the teaching method of the subject area in the learning phase have been described.

<Derived form using major and minor categories>
Hereinafter, a derivation form of the overall configuration of the recognition apparatus according to the present invention will be described.
Here, as a form of derivation of this embodiment, a form in which scenes are roughly divided once by an existing scene classification method, and then a detailed scene is discriminated by applying the present invention will be described.

  Hereinafter, the rough scene class classification is referred to as a large classification scene, and the detailed scene class classification thereafter is referred to as a small classification scene. For example, a major classification scene is a classification such as a party or outdoor sport, and a class such as a Christmas party or a birthday party is considered as a minor classification scene, and soccer or baseball is considered as a sports scene.

As a method for discriminating a large classification scene, for example, Non-Patent Document 6 discloses that a method called a Bag of Words method is effective.
FIG. 13 shows a configuration diagram of the recognition apparatus.
The image input unit 111 receives a camera image. The large classification determination unit 112 classifies party scenes, sports scenes, landscape image scenes, and other various scenes using the Bag of Words method of Non-Patent Document 6.

One of the small classification determination units 113a to 113c is selected based on the determination in the large classification determination unit 112, and the subsequent processing is performed.
The attribute determination unit provided in each of the small classification determination units 113a to 113c includes different parameters (determination dictionaries) learned with different learning data for each large classification.

  For example, in the case of the small classification determination unit 113a that analyzes a party scene, it means that learning is performed by providing only various types of party images as learning data. The candidate region extraction unit and the feature amount extraction unit may also determine the criteria for extracting candidate regions and the types of feature amounts that are suitable for distinguishing small classification scenes.

  These appropriate criteria and feature amounts may be adjusted manually by the user, but a parameter that provides the highest accuracy by giving a combination of a plurality of parameters may be searched by a cross-confirmation method or the like. Thereby, more accurate scene class discrimination specialized for each large classification is performed. The above is the description of the derivative form of the present embodiment. In this way, it has been shown that the present invention can be adapted in a form that can be combined with existing methods.

This is the end of the description of the first embodiment.
As described above, according to the recognition apparatus of the first embodiment, it is possible to analyze an object in an image and discriminate a scene of the image in detail. In addition, scene discrimination learning can be performed without previously teaching what kind of subject should be identified in each scene.
Even if the number of scenes to be discriminated increases, the recognition processing time does not increase significantly. It is also possible to discriminate similar scenes based not only on the presence / absence of a subject but also on the difference in subject variations.

[Second Embodiment]
The second embodiment is an expanded form of the first embodiment. Since this is an extension, only a difference will be briefly described here.
In the first embodiment, the scene class is estimated by classifying an object region as a clue. In this case, there is no clue for discrimination in an image scene in which no object exists, for example, a scene with only a landscape such as the sea or a mountain.

  Therefore, in the second embodiment, a subject area of a type other than an object is also used as an analysis target, candidate areas of each type are extracted, and which scene class is associated with different attribute classifiers. judge. As described above, it is expected that the scene class determination is more robust than the case where only the object region is a clue by analyzing the scene class in a multifaceted manner using a plurality of types of subjects.

  FIG. 14 shows a basic configuration diagram of the present embodiment. The difference between the configuration diagram of the first embodiment and this configuration diagram is that it includes an extraction unit, a feature amount extraction unit, and an attribute determination unit for candidate regions of three different types of subjects. Specifically, the recognition apparatus includes an object candidate region extraction unit 122a that extracts a candidate region of an object, a human body candidate region extraction unit 122b that extracts a human body region, and a background candidate that extracts a background region such as a sky, a mountain, or a floor. A region extraction unit 122c.

  The object candidate area extraction units 122a to 122c extract objects in the same manner as in the first embodiment. The human body candidate region extraction unit extracts a human body candidate region using a known method such as Non-Patent Document 7. The background candidate region extraction unit extracts a region having a high color or texture uniformity and a large area as a background region. Specifically, the image is divided into a plurality of regions by a texture analysis method as in Non-Patent Document 4, and each is extracted as a background candidate region.

  The feature amount extraction units 123a to 123c extract feature amounts that match the type of subject. Here, the feature amount extraction unit 123a of the object and the feature amount extraction unit 123c of the background region extract the same feature amounts as described in the first embodiment. The human body region feature amount extraction unit 123b divides the detected human body region into three parts, the head, upper body, and lower body, in order to focus on the presence or absence of clothes and head coverings. The color SIFT feature values are extracted from these and connected to form a feature vector.

  As described above, each of the attribute determination units 124a to 124c performs learning specialized for subjects of different types. The learning method is the same as the method of the first embodiment, but here, some variation is added.

  As for the background area, it is easy to happen that, for example, it can be known that it is outdoor from a uniform blue sky area, but it is not possible to know whether it is a camp scene or a baseball scene. It is. Therefore, when the background attribute determination unit 124c is learned by the method of the first embodiment, there is a possibility that overlearning is likely to occur.

  Therefore, the following measures are taken here. For example, as one idea, when the depth of the classification tree exceeds a predetermined number, learning is stopped early to prevent over-learning. Another idea is to employ a discriminator that simply divides data randomly without using an information criterion when training the discriminator of each node.

  This is realized by setting the coefficient of the discriminant function of each classifier to a random value. This discriminator is a mathematically similar operation to a known machine learning method called a hash method, random projection, search for similar case data, and approximate nearest neighbor search. That is, the classifier of the attribute determination unit of the present invention is not limited to the ensemble classification tree, and various machine learning methods as described above can be applied.

Next, the determination result integration unit 125 integrates the voting results of the attribute determination units 124a to 124c. At this time, since there is a difference in the reliability of determination obtained depending on the type of subject, the weight W = [w ₁ , w ₂ , w ₃ ] ^T is defined without adding as it is, and a weighted sum is finally obtained. The attribute of the image is identified and used as the final determination result. However, W may be determined by a cross-confirmation method from a plurality of candidate values, or may be determined by solving a value with the best discrimination accuracy on average by the least square method. In addition, instead of simply performing the weighted sum, a form in which the above voting result is input and learning is performed using a discriminator such as a support vector machine so that an image attribute is output may be used.

  The above is the description of the second embodiment in which the scene class is determined using a plurality of types of subjects.

[Third Embodiment]
The recognition apparatus according to the third embodiment is intended to determine a moving image scene class using a moving image. In the present embodiment, a mode for determining whether the monitoring camera image is normal or abnormal will be described. Although there are a plurality of types of abnormal states of surveillance camera images, here, a recognition device that discriminates abnormal behavior of a crowd will be described.

FIG. 15 shows a basic configuration diagram of the recognition apparatus of the present embodiment.
A moving image of the surveillance camera is input to the image input unit 131. The human body candidate region extraction unit 132a is a part that detects a human body by a human body detector technique such as Non-Patent Document 7 as in the second embodiment. The human body extracted here calculates the appearance feature amount in the human body candidate region by the human body feature amount extraction unit 133a.

  Further, the moving direction and moving speed of the human body and the motion feature vector are extracted with reference to the preceding and following moving image frames and added to the feature amount. Here, as in the first embodiment, the visible feature is a color SIFT feature histogram, and the motion feature vector is a feature value of a known method disclosed in Non-Patent Document 8 called CHLAC, for example. .

  Next, the human body attribute determination unit 134a determines the feature amount of the human body region with a discriminator made of an ensemble classification tree, and the candidate region of each human body is a human body of an abnormal scene or a normal scene as in the first embodiment. Get the likelihood score of the human body. Similar to the first embodiment, the human body attribute determination unit 134a extracts the human body from the learning data of the moving image of the abnormal behavior and the moving image of the normal scene in advance by the human body candidate region extraction unit 132a, and classifies the feature amounts of both into the classification tree. Have learned by. Thereby, it is possible to determine a human body region under abnormal behavior such as a person who is swinging a weapon, a person who runs away, or a person who is fighting, as having a high degree of abnormality.

  The human body detection method as in Non-Patent Document 7 has high detection accuracy of a standing human body without shielding, but a person whose head is partially visible in a crowd or a person with a small appearance size However, there is a problem that it is difficult to detect a person in a posture other than standing. FIG. 16A shows an example of a scene of a moving image of a crowd. As illustrated by a black rectangular frame in FIG. 16B, only a small number of human bodies are detected under such conditions. There are many things.

  Therefore, as a contrivance for such a problem, the present recognition apparatus extracts not only individual human bodies but also candidate areas that can be determined by the crowd candidate area extraction unit 132b to be highly likely to have a crowd. Then, whether the same area is a crowd in a normal scene or a crowd in an abnormal scene is determined by examining the feature amount of the area as a clue.

The extraction of the candidate region of the crowd uses the CHLAC feature previously used as the motion feature vector (refer to Non-Patent Document 8 for details). Using this, processing is performed as follows.
First, as the learning data, various crowd videos and videos that do not include the crowd but have various movements are prepared. These moving images are divided into space-time blocks of a certain size such as 16 pixels × 16 pixels × 16 frames, and a CHLAC feature vector is extracted from each block.

  Since the CHLAC feature is a 251-dimensional feature quantity, distribution of two groups of samples equal to the number of blocks is obtained in the 251-dimensional space. Next, linear discriminant analysis, which is a general discriminating method, is performed on the two groups of data to obtain a projection vector onto the best one-dimensional basis for dividing the two groups of data.

  Next, the operation of the crowd candidate area extraction unit 132b at the time of recognition will be described. First, when an input moving image is received, it is divided into block data of the same size as at the time of learning (the blocks may overlap each other). A CHLAC feature is extracted from each block and projected onto a one-dimensional basis obtained by linear discriminant analysis previously to obtain a value on the basis.

  Since the value on the basis is the likelihood of the crowd likeness of each block, it is cut by a predetermined threshold value and extracted as a crowd candidate region. FIG. 16C shows an example of the result with a thick black border. In the subsequent processing, the features of the appearance and movement are extracted by the crowd feature amount extraction unit 133b in the same manner as the previous embodiments, and the discrimination attribute is obtained by determining whether the crowd attribute determination unit 134b is abnormal or normal.

  Next, the region to be extracted by the object candidate region extraction unit 132c is an unspecified subject region that may be observed with abnormal crowd behavior. For example, various subjects such as smoke from a smoke lamp, flames generated by burning objects on the road, scattering of fragments due to vandalism, and the like can be considered.

  Recognizing apparatus of this embodiment obtained by applying the present invention, during pre any particular subject in one to perform the learning features without having any assumptions relating to the determination of the scene classes in the learning is there. Therefore, the object candidate area extraction unit 132c is configured so that various object candidate areas can be extracted here. Specifically, a region in which features of motion and appearance are similar is extracted as an object candidate region.

  The specific method of extracting the object candidate region is an extension of the method using Super-pixel (hereinafter referred to as SP) described with reference to FIG. 3 in the first embodiment. There are two differences from the method of FIG. The first point is to use not only the appearance feature of the pixel but also the motion feature as the similarity when the SP is connected. The second point is to exclude the SP area without movement from the candidate area first. Details will be described below.

  First, for one frame in a moving image, an optical flow, which is a general technique in moving image analysis, is calculated for each pixel. Next, SPs are created. At this time, SPs whose average amount of optical flows in the region is below a certain level are deleted.

  Next, similar SPs similar to S302 to S306 in FIG. 3 are connected, and as the similarity used at this time, both the RGB color distribution of SP and the optical flow orientation distribution are connected. The similarity is calculated for the combined vectors. By doing in this way, the area | region where appearance and motion resembled can be extracted as an object candidate area | region.

FIG. 16D shows an example of the result of the operation of the object candidate region extraction unit 132c with a black thick frame line. Here, a flame and smoke area and a part of the crowd are extracted as object candidate areas.
Note that the method of extracting the object candidate region is not limited to the method described here, and any method can be applied as long as it is a method of extracting a region having a similar appearance and motion.
Subsequent object feature amount extraction unit 133c and object attribute determination unit 134c are the same as those for the previous crowd area. Since it is repeated, detailed description is omitted.

As described above, a score as to whether or not the video of the input moving image is a normal crowd scene is obtained from each candidate region using a plurality of types of candidate regions as clues.
Next, as in the second embodiment, the determination result integration unit 135 votes the discrimination score and totals it for each subject type. Next, the obtained scores are weighted and summed. Furthermore, since the results for each frame may not be stable, the results are moving averaged between a plurality of previous and subsequent frames to obtain the final result.

  In addition to the CHLAC feature, there are various motion vector features such as a method using a spatio-temporal gradient histogram, a method using a hidden Markov model, and so on, and it is only necessary to select one that suits the type of subject.

  As a derivation of the above-described embodiment, when learning each attribute determination unit, for example, if fire motion data and normal scene motion data are given as learning data to learn 2-class scene discrimination, fire detection A recognition device having a function can be realized. Alternatively, learning can be performed by classifying by giving teacher values of classes of three scenes of fire, abnormal behavior, and normal. Thus, the present invention can be applied to various real problems.

In the present embodiment, a moving image is used. However, a derivative form using a distance image in combination is also conceivable in order to increase the accuracy as a surveillance camera. In this case, a device may be used in which feature amounts are extracted and connected from both the luminance image and the distance image, and the discrimination is performed by the discriminator.
Above, description of 3rd Embodiment of the form which discriminate | determines the abnormal state of a crowd and a scene from a moving image is completed.

[Fourth Embodiment]
In the fourth embodiment, a composition class of an image is determined using a still image as an input. In this embodiment, the area of the main subject of the image is estimated simultaneously with the determination of the class of the image composition. According to the disclosure of the present embodiment, the present invention can be used not only for determining information of a single variable such as a scene class but also for estimating a complicated image attribute such as a main subject region by an appropriate device. Is shown.

  In general, various types of image composition classes have been proposed, and the types such as “Hinomaru composition”, “Tripartite composition (which is an approximation of the golden ratio)”, “Diagonal composition”, and “Triangular composition” are known. It has been.

  If the photographic composition can be automatically estimated at the time of taking a picture, it is possible to determine appropriate values for the photographic composition such as the focus position and exposure at the time of taking the picture. It is also possible to easily correct the composition by, for example, teaching the user the frame line of the photographic composition that matches the subject.

  In addition, if the area of the main subject is known, the focus and exposure can be appropriately controlled according to the main subject. In addition, information on the main subject is extremely important information for secondary use of images such as organizing images and creating highlights.

  However, there are some problems with the conventional main subject recognition method. For example, in the case of a method that makes a determination based on a degree of saliency such as a color contrast difference, even if it is an area that is not semantically important, it may be mistakenly recognized as a main subject if the color and brightness contrast with the surroundings is strong. there were.

  For example, when a photo is taken in a room where the light is only bright in the corner of the room, or when the white sky is slightly visible from the gap between the house and the house on the street, it is mistakenly identified as the main subject. There was a thing.

  For example, as another main subject recognition method, there is a method of determining a main subject using object detection such as human body detection or face detection. However, in such a method using object detection, there are various unspecified methods. There was a problem that could not be applied to objects.

  In the present embodiment, the subject constituting the image is extracted, and the attribute determination unit classifies and discriminates, estimates the composition class and main subject region for each subject, and integrates them to automatically discriminate these image attributes. Indicates that is possible.

  FIG. 18A shows composition classes of seven types of images, which are classes to be discriminated in the present embodiment. Each image of the learning data is previously assigned a teacher value corresponding to which of the seven classes. The main subject area of each image is given a teacher value of the main subject by a binary image.

  FIG. 17 shows a basic configuration diagram of the present embodiment. This is a configuration according to the form of the second embodiment. One of the differences from the second embodiment is that the determination result integration unit 145 includes two functional parts, an image composition determination integration unit 145a and a main subject region determination integration unit 145b.

In the following, description will be made focusing on differences from the second embodiment.
In this embodiment, there are three types of subjects to be extracted as candidate areas: object, line segment, and human body. The major difference is that information on a line segment that is considered to have a direct effect in estimating the composition is newly used as a subject.

In order to extract a line segment candidate region, a line segment is extracted from the input image by Hough transform, and only a line segment having a predetermined edge strength and length is left as a candidate line segment.
FIG. 18B shows an example of the extraction result of the line segment candidate region with the symbol 1402b. Further, the line segment feature amount extraction unit 143b extracts a color SIFT feature from a predetermined peripheral range of the candidate line segment, and sets this as a feature amount of appearance of the candidate line segment. In addition to this, the center of gravity position, length, inclination, etc. of the candidate line segments are also calculated and connected to form a feature amount.

  For other object candidate regions and human body candidate regions, feature quantities are extracted by the same method as in the second embodiment. However, in addition to the feature amount used in the second embodiment, the center of gravity position of each region considered to be important in determining the composition and the feature amount of the secondary moment of the region shape are also extracted and added. Furthermore, considering the fact that high-contrast areas and in-focus areas have a great influence on determining composition, the difference in color contrast between the inside and outside of the area, the ratio of the amount of edges inside and outside the area, etc. Feature quantities related to contrast and focus are also added and extracted.

  In the subject attribute determination units 144a to 144c, the ensemble classification tree is learned in advance by associating the feature amount of the subject with the composition class of the image to which the subject belongs. This is similar to the relationship in the second embodiment in which learning is determined by associating the feature amount of each subject candidate area with the scene class. If there is a small amount of information that can be used to determine the composition class, such as the position, size, and boundary strength, which are the feature quantities of each subject, the same as when determining the scene class by integrating them. Thus, the composition class can also be correctly determined.

  FIG. 18B shows a schematic example of the result of estimating the composition class using such an ensemble classification tree. The composition classes are estimated for the individual subject areas of the extracted subject areas 1402a to 1402c, and the voting results are shown in the voting spaces 1403a to 1403c. These are weighted and integrated to obtain a final result 1404. In the example result diagram, a three-part composition is output as the final answer.

  Next, a method for estimating the main subject area will be described. In order to estimate the main subject area at the same time as the composition class, in the present embodiment, when learning the composition class using the ensemble classification tree, the following measures are also taken. That is, not only the composition class ratio of the learning data but also information on the main subject area of the learning data is stored in the leaf node of each classifier. Data other than the learning target variable is called metadata.

  The main subject area information, which is metadata, first normalizes the ratio to an image with an aspect ratio of 1: 1. Next, the main subject area images are averaged in the leaf nodes, and this is used as the prior distribution of the main subjects in the leaf nodes. Furthermore, this is approximated by a Gaussian distribution and only the parameters of the Gaussian distribution are stored. The approximation is performed without using the prior distribution itself because the storage capacity and the speed of the voting calculation process become enormous as the size of the classification tree increases. If there is no problem in this respect, the distribution itself may be stored and used without approximation. The Gaussian distribution may be approximated by a mixed Gaussian distribution or the like.

  An example of the metadata of the main subject area obtained in this way is shown in FIG. Here, the prior distribution (approximate with Gaussian distribution) of the main subject area stored in the leaf node reached when the empty area 1411 extracted as the object candidate area is discriminated by the classification tree is symbolized 1412 is shown.

  In this example, in the case of an image in which a uniform area is shown in the upper part of the image, such as the empty area 1411, the main subject area often does not exist above, so that the center as shown in FIG. This indicates that there is a high possibility that the main subject area is at a position centered around the vicinity.

  FIG. 19B shows a schematic diagram of an example of a result obtained by referencing the metadata of the pre-distribution of the main subject area for each candidate subject and voting it. Here, the result of voting from each subject is indicated by symbols 1413a to 1413c, and the final estimation result after weighted integration is indicated by symbol 1414. This figure shows a state in which the vicinity of the position (person) of the main subject is correctly estimated to some extent.

  This is the end of the description of the fourth embodiment for estimating the composition class and main subject area of the image. It has been shown that composition estimation can be performed by using a subject pattern in an image as a clue by an implementation example to which the present invention is applied. It was also shown that the main subject area can be estimated by learning by using metadata in association with the subject classification result. This is a point that is greatly different from the conventional method in which the degree of saliency is determined by a mechanical standard.

  According to the embodiment of the present invention, various image attributes are discriminated by using the subject in the image as a clue, such as the composition of the image, the normality / abnormality of the behavior of the crowd in the image, and the main subject. It is possible.

(Other embodiments)
The present invention can also be realized by executing the following processing. That is, software (computer program) that implements the functions of the above-described embodiments is supplied to a system or apparatus via a network or various computer-readable storage media. Then, the computer (or CPU, MPU, etc.) of the system or apparatus reads out and executes the program.

101 Image Input Unit 102 Candidate Area Extraction Unit 103 Feature Quantity Extraction Unit 104 Attribute Determination Unit 105 Determination Result Integration Unit

Claims (19)

Candidate area extracting means for extracting a plurality of candidate areas of the subject from the input image;
Feature quantity extraction means for extracting feature quantities of the candidate areas from each of the plurality of candidate areas;
For each of said plurality of candidate areas, the based on the feature amount extracted from the candidate regions, by referring to the learning result of associating the feature amount extracted from the object region and the attribute of the learning image of the learning image Te, an attribute determining means for determining an attribute of the input image to which the candidate region is extracted,
A recognition apparatus comprising: an identification unit that identifies attributes of the input image by aggregating the determination results of the attribute determination unit for the plurality of candidate regions .   The recognition apparatus according to claim 1, wherein the candidate area extraction unit extracts candidate areas of a plurality of subjects based on a criterion of a predetermined candidate area. The candidate region extraction means, recognition apparatus according to claim 2, characterized in that it comprises a plurality of candidate region extracting means based on a plurality of different said reference. Said attribute determining means, recognition apparatus according to claim 3, characterized in that it comprises a plurality of attribute determination means corresponding to a plurality of different said candidate region extracting means.   The candidate area extraction unit extracts the plurality of candidate areas by repeating a process of combining two adjacent superpixels having the highest similarity among the plurality of superpixels generated by dividing the image. The recognition apparatus according to claim 2.   The recognition apparatus according to claim 5, wherein the candidate area extracting unit extracts, as the plurality of candidate areas, a super pixel that is generated by repeating the processing and has a larger area than a predetermined value among the super pixels. .   The recognition apparatus according to claim 1, wherein the candidate area extraction unit includes a human body candidate area extraction unit that extracts a human body candidate area from the image. The attribute determination unit performs the learning by classifying the object region so as to be separated according to the attribute of the image when learning the attribute of the object region of the learning image . 8. The recognition device according to any one of items 7. The recognition apparatus according to claim 8, wherein the attribute determination unit performs learning so as to classify a region taught as an important object region preferentially when the object region is classified.   The recognition apparatus according to claim 1, wherein the attribute determination unit performs learning based on a candidate area of a subject that is not related to the attribute of the image.   The recognition apparatus according to claim 1, wherein the attribute determination unit is configured by a method based on a classification tree.   The recognition apparatus according to claim 1, wherein the attribute determination unit is configured by a method based on a search for similar case data.   The recognition apparatus according to claim 1, wherein the attribute determination unit is configured by a technique based on a hash method.   The attribute to be determined by the attribute determining means is any one of an image scene, a crowd action in the image, a composition type of the image, information on a main subject of the image, and information on a light source direction of the image. The recognition device according to any one of claims 1 to 13.   The recognition apparatus according to claim 1, wherein the image for which the attribute is determined is a moving image.   The recognition apparatus according to claim 1, wherein the image for which the attribute is determined is a distance image. The identification unit excludes the determination result of the attribute determination unit for a candidate region having a high ratio determined not to be an object in the learning image when the determination result of the attribute determination unit for the plurality of candidate regions is totaled. The recognition apparatus according to any one of claims 1 to 16, wherein the recognition apparatus includes: A candidate area extracting step of extracting a plurality of candidate areas of the subject from the input image;
A feature amount extraction step of extracting a feature amount of the candidate region from each of the plurality of candidate regions;
For each of said plurality of candidate areas, the based on the feature amount extracted from the candidate regions, by referring to the learning result of associating the feature amount extracted from the object region and the attribute of the learning image of the learning image An attribute determination step of determining an attribute of the input image from which the candidate area is extracted ;
An identification step of identifying the attributes of the input image by aggregating the determination results of the attribute determination step for the plurality of candidate regions . A candidate area extracting step of extracting a plurality of candidate areas of the subject from the input image;
A feature amount extraction step of extracting a feature amount of the candidate region from each of the plurality of candidate regions;
For each of said plurality of candidate areas, the based on the feature amount extracted from the candidate regions, by referring to the learning result of associating the feature amount extracted from the object region and the attribute of the learning image of the learning image An attribute determination step of determining an attribute of the input image from which the candidate area is extracted ;
A program for causing a computer to execute an identification step of identifying attributes of the input image by counting the determination results of the attribute determination step for the plurality of candidate regions .

Images