JP6210856B2 - Object position specifying system and object position specifying method - Google Patents

Description

  The present invention relates to an object position specifying system and an object position specifying method.

  2. Description of the Related Art Conventionally, there is a technique for recognizing an object shown in an image, that is, a subject (target object) or a scene where an image is captured (see Non-Patent Document 1). In this technology, the input image is processed according to the following procedure, so that the object shown in the image is a data in which a large number of images are grouped and classified according to the type of object (hereinafter referred to as “teacher”). The degree of similarity is calculated for each data), and the object represented by the teacher data with the highest degree of similarity is recognized as the object shown in the input image. To do. That is, the input image is recognized as a scene in which an object having the highest similarity is shown.

(Procedure 1-1): The area of the image is finely divided, and a local feature vector is generated for each divided area.
(Procedure 1-2): A quantization vector is generated for each divided region based on the generated local feature vector.
(Procedure 2): A histogram of the entire image is generated from the generated quantization vector.
(Procedure 3): For example, the generated histogram of the entire image is compared with each of a large amount of teacher data by, for example, SVM (Support vector machine) operation, and the degree of similarity for each classified teacher data Is calculated.

  By the way, the technique disclosed in Non-Patent Document 1 is required to identify the position in the image in which the object is captured after recognizing that the object has the highest similarity. There is a case. In this case, after the scene recognition process is performed on the input image, the scene recognition process is performed again for each area obtained by dividing the image into predetermined areas. It is considered that the position where the object is shown, that is, the position where the object is shown can be specified.

G. Csurka, C.I. R. Dance, Lixin Fan, J.M. Willamowski, C.I. Bray, “Visual Categorization with Bags of Keypoints”, Proc. ECCV Workshop on Statistical Learning in Computer Vision, pp. 59-74, 2004.

  However, after performing the scene recognition process on the input image, performing the scene recognition process on the divided area again means performing the scene recognition process on the number of divided areas + 1 times. There is a problem that the calculation time required for specifying the position in the image in which the object is shown becomes long.

  The present invention has been made on the basis of the above problem recognition, and is capable of specifying a position where the object is reflected in the input image while suppressing an increase in calculation time. An object is to provide a position specifying system and an object position specifying method.

In order to solve the above-described problem, the object position specifying system according to the present invention divides an entire area of an input image into a plurality of first areas having a predetermined first size, and divides the area. For each of the first regions, a local feature vector generation unit that generates a local feature vector that represents a local feature in the image data included in the first region, and a local feature vector generation unit that generates the local feature vector, Quantizing the value of the local feature vector of the first region and generating a quantization vector corresponding to each of the first regions, and each of the quantization vector generation units generated by the quantization vector generation unit A quantization vector storage unit that stores the value of the quantization vector for each of the first regions, and a region of the whole or a part of the image from the values of the quantization vectors for each of the first regions. A histogram generator for generating a histogram representing a and SVM operation unit for performing support vector machine (SVM) operation on said histogram the histogram generating unit has generated, said local feature vector generating unit, and the quantized vector generation unit Each of the histogram generation unit and the SVM calculation unit is controlled to execute a scene recognition process for recognizing the scene of the image in which the object is copied, and then the determination is made in the scene recognition process. For specifying in which position of the plurality of second areas the object is divided into a predetermined second size larger than the first area of the entire area of the image A position specifying control unit that executes a position specifying process, wherein the position specifying control unit includes the histogram in the scene recognition process. A ram generating unit that generates a histogram representing the entire image from the quantization vector value for each of the first regions, and the SVM calculating unit includes a histogram representing the entire image, and a plurality of histograms. An SVM operation is performed to compare each of the histograms of a plurality of teacher data in which the histogram of the image is classified and grouped for each type of object, and in the position specifying process, the histogram generation unit performs the quantization A histogram representing each image of the second region is generated from the value of the quantization vector for each of the first regions stored in the vector storage unit, and the SVM operation unit causes the second operation to be performed. An SVM operation is performed on each histogram representing a region.

  The object position specifying system according to the present invention further includes a histogram storing unit that stores a histogram representing the entire image generated by the histogram generating unit, and the position specifying control unit includes the position specifying process. The SVM calculation unit is caused to execute an SVM calculation for comparing each of the histograms representing the second region with a histogram representing the whole of the image stored in the histogram storage unit. .

  Further, in the position specifying process, the position specifying control unit according to the present invention causes the SVM calculating unit to set each of the histogram representing the second region and a predetermined condition among the plurality of teacher data. A SVM operation for comparing each of the histograms of a part of the selected teacher data is executed.

  The object position specifying system according to the present invention includes a histogram storage unit that stores the histogram that represents the entire image generated by the histogram generation unit, and a histogram that represents the entire image stored in the histogram storage unit. Or a teacher data switching unit that selects and outputs one of the histograms of a part of the teacher data selected according to a predetermined condition among the plurality of teacher data, and In the position specifying process, the control unit causes the SVM calculating unit to control each of the histograms representing the second area and the histogram output from the teacher data switching unit by controlling the teacher data switching unit. The SVM operation for comparing the two is executed.

In the object position specifying method of the present invention, the entire area of the input image is divided into a plurality of first areas having a predetermined first size, and each of the divided first areas is divided. A local feature vector generating unit that generates a local feature vector representing a local feature in the image data included in the first region, and each of the first regions generated by the local feature vector generating unit A quantization vector generation unit that quantizes the value of the local feature vector and generates a quantization vector corresponding to each of the first regions, and a value of each of the quantization vectors generated by the quantization vector generation unit Is generated from a quantization vector storage unit for storing each of the first regions and a value of the quantization vector for each of the first regions. A histogram generation unit for, the SVM computation unit for performing support vector machine (SVM) operation on said histogram the histogram generating unit has generated, said local feature vector generating unit, and the quantized vector generation unit, the histogram generation unit And the SVM calculation unit, and after executing the scene recognition process for recognizing the scene of the image in which the object is copied, the object determined in the scene recognition process is A position specifying process for specifying at which position of the plurality of second areas obtained by dividing the entire area of the image into a predetermined second size larger than the first area; An object position specifying system comprising: a position specifying control unit that executes the position recognition control unit; A procedure for generating a histogram representing the whole of the image from the value of the quantization vector for each of the first regions, and a histogram representing the whole of the image in the SVM calculating unit; And a step of executing an SVM operation for comparing each of a plurality of histograms of a plurality of teacher data in which histograms of a plurality of images are classified and collected for each type of object, and in the position specifying process, the histogram A step of causing the generation unit to generate a histogram representing an image of each of the second regions from the value of each of the quantization vectors stored in the quantization vector storage unit; Including causing the SVM calculation unit to execute SVM calculation for each of the histograms representing the second region. Features.

  According to the present invention, it is possible to specify the position where an object is shown in an input image while suppressing an increase in calculation time.

1 is a block diagram showing a schematic configuration of an object position specifying system according to a first embodiment of the present invention. It is the flowchart which showed the process sequence in the target object location identification system of the 1st embodiment. It is the figure which showed typically an example of the whole process in the target object location specifying system of the 1st embodiment. It is the figure which showed typically an example of the operation | movement which performs the process of scene recognition in the target object location specifying system of the 1st embodiment. It is a figure explaining the view of the process which pinpoints the position of a target object in the target object location specifying system of the 1st embodiment. It is the block diagram which showed schematic structure of the target object location identification system by the 2nd Embodiment of this invention. It is the flowchart which showed the process sequence in the target object location identification system of the 2nd embodiment. It is a figure explaining the view of the process which specifies easily the position of a target object in the target object specifying system of the 2nd embodiment. It is the block diagram which showed schematic structure of the target object location identification system by the 3rd Embodiment of this invention. It is the flowchart which showed the process sequence in the target object location identification system of the 3rd embodiment. It is a figure explaining the view of the process which specifies easily the position of a target object in the target object specifying system of the 3rd embodiment. It is the block diagram which showed schematic structure of the target object location identification system by the 4th Embodiment of this invention.

<First Embodiment>
Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a schematic configuration of an object position specifying system according to the first embodiment. In FIG. 1, an object position specifying system 10 includes a local feature vector generation unit 110, a quantization vector generation unit 120, a histogram generation unit 130, an SVM (Support Vector Machine) operation unit 140, a teacher A data group 150, a position specifying control unit 160, and a quantization vector storage unit 170 are provided.

  The target object specifying system 10 performs scene recognition processing on an input image to recognize an object shown in the image, that is, a subject (target object) and a scene where the image is photographed. Information on the degree of similarity with each teacher data classified for each type of object is output as information determined by scene recognition processing. Further, the object position specifying system 10 performs a position specifying process for specifying a position where the determined object is reflected in the image subjected to the scene recognition process, and determines the position where the specified object is reflected. Output the information that represents it.

  The teacher data group 150 is a database in which histograms of a large number of images showing the same object are included as respective teacher data classified for each type (category) of the object. The teacher data is classified into categories of objects such as people, dogs, cats, and flowers, for example, and is composed of, for example, a histogram of 1500 images for each classified category. That is, in the teacher data group 150, there are 1500 histograms as teacher data for one category whose object is “person”, and similarly, the objects are “dog”, “cat”, For each category of “flower”, 1500 histograms exist as teacher data. That is, the teacher data group 150 includes 1500 histograms (4 × 1500 = 6000 histograms in total) for each of the four categories as teacher data.

  The local feature vector generation unit 110 generates a local feature vector of the image input to the object position specifying system 10 in accordance with the control from the position specifying control unit 160. The local feature vector generation unit 110 finely divides the entire area of the input image into areas of a predetermined size (hereinafter referred to as “scene recognition divided areas”) and includes the divided scene recognition divided areas. Local feature vectors representing local features in the image data to be generated are generated. Then, the local feature vector generation unit 110 outputs the value of the generated local feature vector of each scene recognition divided region to the quantization vector generation unit 120. In addition, when the generation of local feature vectors for all scene recognition divided regions is completed, the local feature vector generation unit 110 notifies the position specification control unit 160 that the generation of the local feature vectors is completed. The local feature vector generation unit 110 generates a local feature vector in the same manner as the processing method for generating a local feature vector when performing scene recognition processing in the prior art. Is omitted.

  The quantization vector generation unit 120 quantizes the local feature vector values of the respective scene recognition divided regions input from the local feature vector generation unit 110 in accordance with the control from the position specifying control unit 160, and each scene recognition. A quantization vector corresponding to the divided region is generated. Then, the quantization vector generation unit 120 outputs the generated quantization vector value of each scene recognition divided region to the histogram generation unit 130, and outputs the generated quantization vector value of each scene recognition divided region. The quantization vector storage unit 170 stores the result. Further, when the generation of the quantization vectors of all scene recognition divided regions is completed, the quantization vector generation unit 120 notifies the position specification control unit 160 that the generation of the quantization vectors is completed. Note that the processing method for generating the quantization vector in the quantization vector generation unit 120 is the same as the processing method for generating the quantization vector when performing the scene recognition processing in the conventional technique, and thus detailed description will be given. Is omitted.

  The quantization vector storage unit 170 temporarily stores the quantization vector value of each scene recognition divided region generated by the quantization vector generation unit 120 according to the control from the quantization vector generation unit 120, for example, And a memory such as a DRAM (Dynamic Random Access Memory). The quantization vector storage unit 170 stores the value of the quantization vector corresponding to each scene recognition divided region input from the quantization vector generation unit 120 for each scene recognition divided region. The value of the quantization vector for each scene recognition divided region stored in the quantization vector storage unit 170 is output to the histogram generation unit 130 in accordance with control from the histogram generation unit 130.

  In the scene recognition process according to the control from the position specifying control unit 160, the histogram generation unit 130 calculates the object from the quantization vector value for each scene recognition divided region input from the quantization vector generation unit 120. A histogram representing the entire image input to the position specifying system 10 is generated. Then, the histogram generation unit 130 outputs the generated histogram of the entire image to the SVM calculation unit 140. Further, when the generation of the histogram of the entire image corresponding to the input image is completed, the histogram generation unit 130 notifies the position specification control unit 160 that the generation of the histogram of the entire image is completed. Note that the processing method for generating the histogram of the entire image in the histogram generation unit 130 is the same as the processing method for generating the histogram of the entire image when performing the scene recognition processing in the conventional technique, and thus detailed description will be given. Is omitted.

  Further, the histogram generation unit 130 performs quantization vector for each scene recognition divided region stored in the quantization vector storage unit 170 in the process of specifying the position of the object according to the control from the position specification control unit 160. Based on the value, a histogram of an area having a predetermined size (hereinafter referred to as “position-specific divided area”) larger than the size of the scene recognition divided area designated by the position specifying control unit 160 is generated. The position specifying divided area is an area in which a unit for specifying a position where the object is shown in the entire image input to the object position specifying system 10 is determined. Then, the histogram generation unit 130 outputs the generated histogram for each position specifying divided region (hereinafter referred to as “position specifying histogram”) to the SVM calculation unit 140. In addition, when the generation of the position specifying histogram corresponding to the position specifying divided region specified by the position specifying control unit 160 is completed, the histogram generating unit 130 determines that the generation of the specified position specifying histogram has been completed. To the unit 160. By this notification, the next position specifying divided region is designated by the position specifying control unit 160, and the histogram generating unit 130 repeats generation of the position specifying histogram corresponding to the designated position specifying divided region. Note that the processing method for generating the position specifying histogram in the histogram generation unit 130 is the same as the processing method for generating the histogram of the entire image in the scene recognition processing, except that the size of the region for generating the histogram is different. .

  In the scene recognition process in accordance with the control from the position specifying control unit 160, the SVM calculation unit 140 includes the histogram of the entire image input from the histogram generation unit 130 and the histogram of each teacher data included in the teacher data group 150. SVM calculation is performed, and similarity is calculated for each category of objects classified in the teacher data group 150. The SVM calculation unit 140 then calculates each of the targets calculated by the SVM calculation when the calculation of the similarity with the category of each target with respect to the entire histogram of the input image is completed, that is, when the SVM calculation is completed. Information indicating the similarity for each category of the object is output as information determined by the object position specifying system 10 performing scene recognition processing. In addition, the SVM calculation unit 140 notifies the position specifying control unit 160 that the SVM calculation for performing the scene recognition processing is completed. Note that the SVM calculation method in the SVM calculation unit 140 is the same as the SVM calculation method used when performing scene recognition processing in the conventional technique, and thus detailed description thereof is omitted.

  In addition, the SVM calculation unit 140 is included in each position specifying histogram input from the histogram generation unit 130 and the teacher data group 150 in the process of specifying the position of the object according to the control from the position specifying control unit 160. An SVM operation for comparing the histograms of the respective teacher data is performed, and the similarity with the category of the object classified in the teacher data group 150 is calculated for each position specifying divided region. In addition, the SVM calculating unit 140 notifies the position specifying control unit 160 that the SVM calculation is completed for each position specifying divided area where the process of specifying the position of the object is performed. By this notification, the next position specifying divided region is designated from the position specifying control unit 160, and the SVM calculating unit 140 repeats the SVM calculation for the position specifying histogram of the specified position specifying divided region. When the SVM calculation unit 140 completes the SVM calculation with respect to the position specifying histograms of all the position specifying divided regions, the SVM calculating unit 140 determines the object identified by the scene recognition process for each position specifying divided region calculated by the SVM calculation. Information representing the similarity to the category is output as a result of the object position specifying system 10 performing the object position specifying process. Note that the SVM calculation method for the position specification histogram in the SVM calculation unit 140 is different from the histogram for performing the SVM calculation process, that is, the SVM calculation in the scene recognition process except that the histogram of the entire image becomes the position specification histogram. It is the same as the method.

  The position specifying control unit 160 is the entire object position specifying system 10, that is, the local feature vector generating unit 110, the quantized vector generating unit 120, the histogram generating unit 130, and the SVM calculating unit included in the object position specifying system 10. Each operation of 140 is controlled. The position specifying control unit 160 includes a histogram generation divided region specifying unit 161.

  The histogram generation divided region specifying unit 161 shows the entire region of the image input to the object position specifying system 10 in the image in the object position specifying process in the object position specifying system 10. It divides | segments into the position specific division area of the predetermined magnitude | size for performing the position specific process which specifies the position which exists. Then, the histogram generation divided region specifying unit 161 sequentially specifies each of the divided position specifying divided regions as a region where the histogram generating unit 130 generates the position specifying histogram and a region where the SVM calculating unit 140 performs the SVM calculation. At this time, the histogram generation divided region designation unit 161 designates the histogram generation unit 130 every time it receives a notification from the histogram generation unit 130 that the generation of the position specific histogram corresponding to the position specific division region designated this time is completed. The position specifying divided area to be moved is sequentially moved to the next position specifying divided area. In addition, the histogram generation divided region designation unit 161 designates a position to be designated to the SVM calculation unit 140 every time it receives a notification from the SVM calculation unit 140 indicating that the SVM calculation for the position specification histogram of the position specific division region designated this time has been completed. The specific divided area is sequentially moved to the next position specific divided area. That is, the histogram generation divided region specifying unit 161 sequentially changes the position specifying histogram for performing the SVM calculation unit to the position specifying histogram of the next position specifying divided region.

  With such a configuration, the target object specifying system 10 stores the quantization vector value of each scene recognition divided region generated by the quantization vector generation unit 120 in the scene recognition processing in the quantization vector storage unit 170. Then, by using the quantization vector value of each scene recognition divided region stored in the quantization vector storage unit 170, a position specifying histogram for each position specifying divided region when performing the process of specifying the position of the object is generated. Thereby, in the object position specifying system 10, even when it is required to specify the position in the image in which the determined object is shown after the scene recognition process, the object position specifying region 10 is determined. The processing for specifying the position of the object can be performed with less processing than performing the processing for scene recognition again.

  Next, the operation of the object position specifying system 10 will be described. FIG. 2 is a flowchart showing a processing procedure in the object position specifying system 10 of the first embodiment. Moreover, FIGS. 3-5 is a figure explaining an example of each process in the target object location specifying system 10 of the 1st embodiment. In the description of the flowchart of the process in the object position specifying system 10 shown in FIG. 2, an example of each process in the object position specifying system 10 shown in FIGS. An example in which the target object is a “dog” will be described.

  When an image is input to the object position specifying system 10, the position specifying control unit 160 first performs a scene recognition process on the input image, and then performs an object position specifying process. The operation of each component provided in the object location specifying system 10 is controlled. FIG. 3 is a diagram schematically illustrating an example of the entire process in the object position specifying system 10 according to the first embodiment.

  When an image in which “dog” is shown at the position shown in FIG. 3A is input to the object position specifying system 10, first, the position specifying control unit 160 first sets the object position specifying system 10 to the object position specifying system 10. By controlling the operation of each of the constituent elements provided, scene recognition processing is performed on the input image (FIG. 3A), and it is determined that “dog” appears in the input image. . Thereafter, the position specifying control unit 160 controls the operation of each component included in the object position specifying system 10, and the entire area of the input image (FIG. 3A) is shown in FIG. 3B. As shown in the figure, the position of the position-specific divided area where the object “dog” is reflected is performed by sequentially performing the process of specifying the position of the object for each position-specific divided area divided into a plurality of position-specific divided areas. Is identified. In FIG. 3B, the entire area of the image (FIG. 3A) is divided into nine position-specific divided areas A1 to A9 that are each divided into three in the horizontal direction and the vertical direction, and the position of the position-specific divided area A6. Is identified as the position where the object “dog” is shown.

  When an image is input to the object position specifying system 10, the object position specifying system 10 starts scene recognition processing for recognizing the scene of the input image from step S100. In the process of scene recognition in the object position specifying system 10, first, in step S 100, the position specifying control unit 160 stores the entire region of the input image (FIG. 3A) in advance in the local feature vector generation unit 110. A local feature vector is generated for each scene recognition divided area divided into small scene recognition divided areas of a predetermined size.

  Subsequently, in step S110, the position specifying control unit 160 causes the quantization vector generation unit 120 to obtain the quantization vector for each scene recognition divided region based on the local feature vector generated by the local feature vector generation unit 110. Generate. In step S115, the position specifying control unit 160 causes the quantization vector generation unit 120 to store the generated quantization vector values of the respective scene recognition divided regions in the quantization vector storage unit 170.

  Subsequently, in step S120, the position specifying control unit 160 causes the histogram generating unit 130 to specify the object position based on the value of the quantization vector for each scene recognition divided region generated by the quantization vector generating unit 120. A histogram representing the entire image (FIG. 3A) input to the system 10 is generated.

  Subsequently, in step S130, the position specification control unit 160 causes the SVM calculation unit 140 to display the entire histogram generated by the histogram generation unit 130 (FIG. 3A) and each teacher included in the teacher data group 150. The SVM calculation for calculating the similarity with the data histogram is executed. Thereby, the object position specifying system 10 can determine that “dog” is reflected in the input image (FIG. 3A), and information indicating the similarity for each category of each object. Is output.

  The process so far is the scene recognition process in the object position specifying system 10. The scene recognition process in the object position specifying system 10 is the same as the scene recognition process according to the conventional technique.

  Here, an example of the scene recognition process from step S100 to step S130 performed by the object position specifying system 10 will be described. FIG. 4 is a diagram schematically illustrating an example of an operation for performing scene recognition processing in the object position specifying system 10 according to the first embodiment.

  In the scene recognition processing in the object position specifying system 10, first, in step S100, the local feature vector generation unit 110 receives an input image (FIG. 3A) in accordance with control from the position specifying control unit 160. The entire area is divided into small scene recognition divided areas having a predetermined size, and a local feature vector is generated for each divided scene recognition divided area. Then, the local feature vector generation unit 110 outputs the value of the generated local feature vector of each scene recognition divided region to the quantization vector generation unit 120. FIG. 4A shows an example of a state of a scene recognition divided region in which the local feature vector generation unit 110 divides the input image into nine parts in the horizontal direction and the vertical direction, respectively.

  In step S110, the quantization vector generation unit 120 quantizes the value of the local feature vector generated by the local feature vector generation unit 110 in accordance with the control from the position specifying control unit 160, and each scene recognition divided region. Generate a quantization vector for each. Then, the quantization vector generation unit 120 outputs the generated quantization vector value of each scene recognition divided region to the histogram generation unit 130. In step S115, the quantization vector generation unit 120 stores the generated quantization vector value of each scene recognition divided region in the quantization vector storage unit 170. FIG. 4B illustrates an example of a state where the quantization vector generation unit 120 has completed the generation of the quantization vector.

  Thereafter, in step S120, the histogram generation unit 130 determines the target object from the quantization vector value for each scene recognition divided region generated by the quantization vector generation unit 120 according to the control from the position specifying control unit 160. A histogram representing the entire image (FIG. 3A) input to the position specifying system 10 is generated. Then, the histogram generation unit 130 outputs the entire histogram of the generated image (FIG. 3A) to the SVM calculation unit 140. FIG. 4C shows an example of the histogram of the entire image generated by the histogram generation unit 130.

  Thereafter, in step S 130, the SVM calculation unit 140 is included in the entire histogram generated by the histogram generation unit 130 (FIG. 3A) and the teacher data group 150 in accordance with the control from the position specifying control unit 160. SVM calculation is performed to calculate the degree of similarity with the histogram of each teacher data (for example, a total of 6000 histograms of the 1500 histograms included in the four categories described above). And the SVM calculating part 140 outputs the information showing the similarity for every category of each target object obtained based on the result of the calculated SVM calculation.

  In the SVM calculation by the SVM calculation unit 140, the absolute difference value of the frequency between the same classes in the histogram of the entire image (FIG. 3A) generated by the histogram generation unit 130 and the histogram of each teacher data is calculated. , The difference absolute value of each class is added. The difference absolute value addition result calculated here indicates that the smaller the difference between the histogram of the entire image (FIG. 3A) generated by the histogram generation unit 130 and the histogram of each teacher data, that is, for each histogram. The higher the similarity, the smaller the value. As a result, the category including the teacher data having the smallest value of the absolute difference value calculated by the SVM calculation unit 140 is the category of the object shown in the input image (FIG. 3A). Can be determined. The SVM calculation unit 140 outputs information indicating the similarity according to the addition result of the absolute difference value. In FIG. 4D, information that the similarity to the category whose object is “Dog” is 80% and information that the similarity to the category whose object is “Cat” is 20% are output. An example of the case is shown.

  Note that if the size of the image area represented by each histogram is different, that is, if the total number of image data (number of pixels) included in the image is different, even if the histograms represent the same image, The frequencies in the same class are different, and the absolute difference value calculated from the frequencies of the same class becomes a large value. For this reason, in the SVM calculation, the number of frequencies included in each histogram is set so that the size of the image area represented by each histogram is equal, that is, the total number of image data is equal. After normalization, the difference absolute value of the frequency between the same classes in each histogram is calculated. The same applies to scene recognition processing according to the conventional technique.

  When the scene recognition process is completed, the object position specifying system 10 performs a position specifying process for specifying the position where the determined object is shown in the image subjected to the scene recognition process from step S200. Start. In the object position specifying process in the object position specifying system 10, the quantization vector value of each scene recognition divided region stored in the quantization vector storing unit 170 by the quantization vector generating unit 120 in the scene recognition process is used. The same processing as the processing after the generation of the histogram of the entire image (step S120) in the scene recognition processing is performed for each position-specific divided region.

  First, in step S200, the position specifying control unit 160 causes the histogram generating unit 130 to divide the entire area of the input image (FIG. 3A) into nine position specifying divided regions (see FIG. 3B). The first position specifying divided area is designated. Then, the position specification control unit 160 causes the histogram generation unit 130 to acquire the quantization vector value corresponding to the designated first position specification divided region from the quantization vector storage unit 170.

  Subsequently, in step S210, the position specifying control unit 160 causes the histogram generating unit 130 to determine the first position specifying divided region based on the value of the quantization vector corresponding to the acquired first position specifying divided region. A location-specific histogram representing is generated.

  Subsequently, in step S220, the position specifying control unit 160 includes the SVM calculation unit 140 in the position specifying histogram representing the first position specifying divided region generated by the histogram generating unit 130 and the teacher data group 150. The SVM calculation for calculating the similarity with the histogram of each teacher data of the category of the object with the highest similarity determined in the scene recognition process is executed. Thereby, the object position specifying system 10 can determine whether or not the “dog” having the highest similarity in the scene recognition processing is included in the first position specifying divided region. Outputs information indicating.

  The process so far is the position specifying process for one position specifying divided region in the object position specifying system 10. In the object position specifying system 10, nine position specifying divided areas (see FIG. 3B) obtained by dividing the entire area of the input image (FIG. 3A), that is, all the position specifying divided areas are obtained. Thus, the position specifying process from step S200 to step S220 is performed.

  More specifically, in step S230, the position specifying control unit 160 determines whether or not the specification of all the position specifying divided areas obtained by dividing the entire area of the input image (FIG. 3A) has been completed. To do. As a result of the determination in step S230, when the designation of all divided position specifying divided areas is completed, the processing in the object position specifying system 10 is completed. On the other hand, if the result of determination in step S230 is that the designation of all the divided position-specific divided areas has not been completed, the process returns to step S200 to designate the next position-specific divided area, and steps S200 to S220. Repeat the position identification process up to.

  Here, the position specifying process from step S200 to step S220 performed by the object position specifying system 10 will be described. FIG. 5 is a diagram for explaining the concept of processing for specifying the position of an object in the object position specifying system 10 of the first embodiment.

  In the process of specifying the position of the object in the object position specifying system 10, first, in step S200, the position specifying control unit 160 divides the entire area of the input image (FIG. 3A) into nine positions specified. Among the divided areas (see FIG. 3B), the first position specifying divided area A1 is designated. Then, the histogram generation unit 130 acquires the quantization vector value corresponding to the first position specifying divided region A1 specified by the position specifying control unit 160 from the quantized vector storage unit 170. In FIG. 5A-1, the histogram generation unit 130 indicates the quantization vector value of each scene recognition divided region corresponding to the first position specifying divided region A1 specified by the position specifying control unit 160. 2 shows an example of a state acquired from the quantized vector storage unit 170.

  After that, in step S210, the histogram generation unit 130 controls the value of the quantization vector corresponding to the first position specifying divided region A1 acquired from the quantization vector storage unit 170 in accordance with the control from the position specifying control unit 160. Then, a position specifying histogram representing the first position specifying divided area A1 is generated. Then, the histogram generation unit 130 outputs a position specifying histogram representing the generated first position specifying divided region A1 to the SVM calculation unit 140. FIG. 5A-2 illustrates an example of a position specifying histogram generated by the histogram generation unit 130 from the quantization vector value corresponding to the first position specifying divided region A1 acquired from the quantization vector storage unit 170. Is shown.

  Thereafter, in step S220, the SVM calculating unit 140, in response to control from the position specifying control unit 160, a position specifying histogram representing the first position specifying divided area A1 generated by the histogram generating unit 130, and a teacher data group 150, the SVM calculation is performed to calculate the similarity with the histogram of each teacher data of the category “dog” having the highest similarity in the scene recognition process. Then, the SVM calculation unit 140 outputs information representing the similarity with the category “dog” obtained based on the result of the SVM calculation calculated for the position specifying divided region A1.

  Note that, in the process of specifying the position of the target object, as in the SVM calculation by the SVM calculation unit 140 in the scene recognition process, the position specifying histogram representing the first position specifying divided region A1 generated by the histogram generating unit 130 The absolute difference value of the frequency between the same classes in the histogram of each teacher data is calculated, and the absolute difference value of each class is added. As a result, the SVM calculation unit 140 determines that the position-specific divided area having the smallest value of the calculated difference absolute value addition value is the position-specific divided area in which the object “dog” is reflected, and Information specifying the position of the position specifying divided area is output.

  Further, when the generation of the position specifying histogram representing the first position specifying divided area A1 designated by the position specifying control section 160 is completed, the histogram generating unit 130 indicates the position indicating the first position specifying divided area A1. The position identification control unit 160 is notified that the generation of the specific histogram has been completed. In response to this notification, the position specifying control unit 160 performs the determination in step S230, returns to step S200, and designates the second position specifying divided region A2.

  Then, the histogram generation unit 130 acquires the quantization vector value corresponding to the second position specifying divided region A2 specified by the position specifying control unit 160 from the quantized vector storage unit 170. In FIG. 5B-1, the histogram generation unit 130 indicates the quantization vector value of each scene recognition divided region corresponding to the second position specifying divided region A2 specified by the position specifying control unit 160. 2 shows an example of a state acquired from the quantized vector storage unit 170.

  Thereafter, in step S210, the histogram generation unit 130 controls the value of the quantization vector corresponding to the second position specifying divided region A2 acquired from the quantization vector storage unit 170 in accordance with the control from the position specifying control unit 160. Thus, a position specifying histogram representing the second position specifying divided area A2 is generated. Then, the histogram generation unit 130 outputs a position specifying histogram representing the generated second position specifying divided region A2 to the SVM calculation unit 140. FIG. 5B-2 shows an example of the position specifying histogram generated from the value of the quantization vector corresponding to the second position specifying divided region A2 acquired by the histogram generation unit 130 from the quantization vector storage unit 170. Is shown.

  Thereafter, in step S220, the SVM calculation unit 140, in response to control from the position specification control unit 160, a position specifying histogram representing the second position specifying divided region A2 generated by the histogram generation unit 130, and a teacher data group 150, the SVM calculation is performed to calculate the similarity with the histogram of each teacher data of the category “dog” having the highest similarity in the scene recognition process. Then, the SVM calculation unit 140 outputs information representing the similarity to the category “dog” obtained based on the result of the SVM calculation calculated for the position specifying divided region A2.

  Thereafter, similarly, the position specifying control unit 160 sequentially designates nine position specifying divided regions obtained by dividing the entire region of the input image (FIG. 3A), and the histogram generating unit 130 determines the position specifying control unit. A position specifying histogram representing each position specifying divided region designated by 160 is sequentially generated and output to the SVM calculation unit 140. Similarly, the SVM calculation unit 140 has the highest similarity in the scene recognition process included in the teacher data group 150 and the location specification histogram representing each location specification divided region generated by the histogram generation unit 130. An SVM calculation is performed to calculate the similarity to the histogram of each teacher data of the category “Dog”. Then, the SVM calculation unit 140 outputs information representing the degree of similarity with the category “dog” in each position specifying divided region, which is obtained based on the result of the calculated SVM calculation.

  Further, the SVM calculation unit 140 selects the position-specific divided area having the highest similarity as the object “dog” based on the information indicating the similarity to the category “dog” in each position-specific divided area. Is determined to be a position-specific divided area, and information for specifying the position of the position-specific divided area is output.

  In this manner, the object position specifying system 10 repeats the position specifying process for all the position specifying divided areas obtained by dividing the entire area of the input image (FIG. 3A), thereby specifying each position. Identifying the position-specific divided area having the highest similarity to the category “dog” among the divided areas as the position-specific divided area in which the “dog” determined by the scene recognition process is reflected as an object Can do. Thereby, the target object specifying system 10 can output information representing the position of the specified position specifying divided region.

  According to the present embodiment, the entire area of the input image is divided into a plurality of first areas (scene recognition divided areas) having a predetermined first size, and each divided scene recognition divided area is divided. In addition, a local feature vector generation unit (local feature vector generation unit 110) that generates a local feature vector representing a local feature in the image data included in the scene recognition divided region, and a local feature vector generation unit 110 generate, A quantization vector generation unit (quantization vector generation unit 120) that quantizes local feature vector values of each scene recognition division region and generates a quantization vector corresponding to each scene recognition division region; Each quantization vector value generated by the unit 120 is stored for each scene recognition divided area (quantized vector storage). 170) and a histogram generation unit (histogram generation unit) that generates a histogram (a histogram of the entire image or a position specifying histogram) that represents the entire image or a partial region of the image from each quantization vector value for each scene recognition divided region 130), an SVM calculation unit (SVM calculation unit 140) that performs a support vector machine (SVM) calculation on the histogram generated by the histogram generation unit 130, and each of the components in the object position specifying system 10 are controlled. After executing the scene recognition process to recognize the scene of the image in which the object is copied, the object (for example, “dog”) determined in the scene recognition process divides the entire area of the image into the scene recognition division A plurality of second areas (position-specific divisions) divided into a predetermined second size larger than the area A position specifying control unit (position specifying control unit 160) that executes a position specifying process for specifying at which position in the area), the position specifying control unit 160 performs scene recognition processing. The histogram generation unit 130 generates a histogram representing the entire image from the quantization vector value for each scene recognition divided region, and the SVM calculation unit 140 includes a histogram representing the entire image and a plurality of images. The histogram generation unit 130 causes the histogram generation unit 130 to perform a quantization vector storage unit in a position specifying process by executing an SVM operation for comparing the histograms of the plurality of teacher data with each of the histograms of the plurality of teacher data collected and classified for each type of object. From the value of each quantization vector for each scene recognition divided area stored in 170, the image of each position-specific divided area is displayed. An object position specifying system (object position specifying system 10) is configured to generate a histogram representing an image and cause the SVM calculating unit 140 to execute an SVM calculation for each of the histograms indicating the position specifying divided regions.

  Further, according to the present embodiment, the entire area of the input image is divided into a plurality of first areas (scene recognition divided areas) having a predetermined first size, and the divided scene recognition divided areas are divided. For each region, a local feature vector generation unit (local feature vector generation unit 110) that generates local feature vectors representing local features in the image data included in the scene recognition divided region, and a local feature vector generation unit 110 generate A quantization vector generation unit (quantization vector generation unit 120) that quantizes local feature vector values of each scene recognition divided region and generates a quantization vector corresponding to each scene recognition divided region; Each quantization vector value generated by the vector generation unit 120 is stored for each scene recognition divided region (quantization vector vector). A storage unit 170) and a histogram generation unit (histogram) that generates a histogram (a histogram of the entire image or a position specifying histogram) that represents the whole or a partial region of the image from each quantization vector value for each scene recognition divided region. Generation unit 130), an SVM calculation unit (SVM calculation unit 140) that performs a support vector machine (SVM) calculation on the histogram generated by the histogram generation unit 130, and each of the components in the object position specifying system 10. After executing the scene recognition process for recognizing the scene of the image in which the object is copied, the object (for example, “dog”) determined in the scene recognition process is used for the entire area of the image. A plurality of second areas (position characteristics) divided into a predetermined second size larger than the recognition divided area. In the object position specifying system, the position specifying control unit includes a position specifying control unit (position specifying control unit 160) that executes a position specifying process for specifying at which position in the divided area). In the scene recognition process, the unit 160 causes the histogram generation unit 130 to generate a histogram representing the entire image from the quantization vector value for each scene recognition divided region, and the SVM calculation unit 140 And a procedure for executing an SVM operation for comparing each of a plurality of histograms of teacher data in which histograms of a plurality of images are classified and grouped for each type of object. In this process, the histogram generation unit 130 stores each of the scene recognition divided regions stored in the quantization vector storage unit 170. A procedure for generating a histogram representing an image of each position-specific divided region from each quantization vector value, and a procedure for causing the SVM calculation unit 140 to perform an SVM calculation on each of the histograms representing the position-specific divided regions The object position specifying method is configured.

  As described above, in the target object specifying system 10 of the first embodiment, the quantization vector value of each scene recognition divided region generated by the quantization vector generation unit 120 in the scene recognition process is set as follows. The data is stored in the quantized vector storage unit 170. In the object position specifying process in the object position specifying system 10 according to the first embodiment, the quantization vector value of each scene recognition divided region stored in the quantization vector storage unit 170 is used. A position specifying histogram for each position specifying divided region is generated. Thereby, in the object position specifying system 10 of the first embodiment, after performing the scene recognition process on the input image in order to specify the position in the image in which the object is shown, The position specifying process can be performed with less processing than performing the same process as the scene recognition process again for each position specifying divided region. That is, in the object position specifying process in the object position specifying system 10 of the first embodiment, a process of generating a local feature vector in the scene recognition process (step S100) and a process of generating a quantization vector (Step S110) can be omitted. Thereby, in the object position specifying system 10 of the first embodiment, it is possible to shorten the calculation time required for specifying the position in the image in which the object is shown.

  In the target object specifying system 10 according to the first embodiment, the SVM calculation unit 140 uses the similarity of each target object category obtained based on the calculated SVM calculation result in the scene recognition process. In the above description, the configuration is described in which the information indicating the position of the position-specific divided region in which the determined target object is captured is output in the process of specifying the position of the target object. However, the component that outputs the information indicating the degree of similarity for each category of the object and the information for specifying the position of the position-specific divided area where the determined object is reflected is not limited to the SVM calculation unit 140. Absent. For example, based on the result of each SVM calculation calculated by the SVM calculation unit 140, the position specifying control unit 160 shows information indicating the similarity for each category of each target object, or the position where the determined target object is reflected. A configuration for outputting information for specifying the position of the specific divided region may be employed.

<Second Embodiment>
Next, a second embodiment of the present invention will be described. FIG. 6 is a block diagram showing a schematic configuration of the object position specifying system according to the second embodiment. 6, the object position specifying system 20 includes a local feature vector generation unit 110, a quantization vector generation unit 120, a histogram generation unit 230, an SVM calculation unit 240, a teacher data group 150, and a position specification control unit. 260, a quantization vector storage unit 170, and a histogram storage unit 280.

  The object position specifying system 20 shown in FIG. 6 includes the histogram generating unit 130 included in the object position specifying system 10 of the first embodiment shown in FIG. Is replaced with the SVM calculation unit 240 and the position specifying control unit 160 is replaced with the position specifying control unit 260, respectively, and further includes a histogram storage unit 280. Further, the other components provided in the object position specifying system 20 are the same as the components provided in the object position specifying system 10 of the first embodiment shown in FIG. Therefore, in the description of the object position specifying system 20 according to the second embodiment, only components and operations different from those provided in the object position specifying system 10 according to the first embodiment will be described. Detailed description of the same components and operations as those of the object position specifying system 10 of the embodiment will be omitted.

  Similar to the target object specifying system 10 of the first embodiment, the target object specifying system 20 is configured to display a subject (target object) in the image and a scene in which the image is captured with respect to the input image. Recognizing scene recognition processing is performed, and information on the degree of similarity with each teacher data classified for each type of object is output as information determined by scene recognition processing. In addition, the object position specifying system 20 is a position for specifying the position where the determined object is shown in the image subjected to the scene recognition process, similarly to the object position specifying system 10 of the first embodiment. A specific process is performed, and information indicating the position where the specified object is shown is output.

  The local feature vector generation unit 110 generates a local feature vector of the image input to the object position specifying system 20 for each scene recognition divided region in accordance with the control from the position specifying control unit 260, and generates each of the generated scenes. The value of the local feature vector of the recognition divided region is output to the quantized vector generation unit 120.

  The quantization vector generation unit 120 quantizes the local feature vector value of each scene recognition division region input from the local feature vector generation unit 110 in accordance with the control from the position specifying control unit 260. Each quantization vector is generated, and the generated quantization vector value of each scene recognition divided region is output to the histogram generation unit 230 and stored in the quantization vector storage unit 170.

  The quantization vector storage unit 170 stores the value of the quantization vector input from the quantization vector generation unit 120 and corresponding to each scene recognition divided region for each scene recognition divided region. The value of the quantization vector for each scene recognition divided region is output to the histogram generation unit 230 in accordance with the control from the histogram generation unit 230.

  Similar to the histogram generation unit 130 provided in the object position specifying system 10 of the first embodiment, the histogram generation unit 230 generates a quantization vector in the process of scene recognition according to the control from the position specification control unit 260. A histogram representing the entire image input to the object position specifying system 20 is generated from the value of the quantization vector for each scene recognition divided region input from the unit 120. Then, the histogram generation unit 230 outputs the generated histogram of the entire image to the SVM calculation unit 240. Further, unlike the histogram generation unit 130 provided in the object position specifying system 10 of the first embodiment, the histogram generation unit 230 causes the histogram storage unit 280 to store the generated histogram of the entire image. Further, when the generation of the histogram of the entire image corresponding to the input image is completed, the histogram generation unit 230 notifies the position specification control unit 260 that the generation of the histogram of the entire image is completed. The processing method for generating the histogram of the entire image in the histogram generation unit 230 is the same as the histogram generation unit 130 provided in the object position specifying system 10 of the first embodiment. Since this is the same as the method of generating a histogram of the entire image when performing the above, detailed description thereof will be omitted.

  In addition, the histogram generation unit 230 performs the process of specifying the position of the object in accordance with the control from the position specifying control unit 260, similarly to the histogram generation unit 130 provided in the object position specifying system 10 of the first embodiment. From the quantization vector value for each scene recognition divided region stored in the quantized vector storage unit 170, a position specifying histogram representing the position specifying divided region designated by the position specifying control unit 260 is generated. Then, the histogram generation unit 230 outputs the generated position specifying histogram for each position specifying divided region to the SVM calculating unit 240. In addition, when the generation of the position specifying histogram corresponding to the position specifying divided region specified by the position specifying control unit 260 is completed, the histogram generating unit 230 determines that the generation of the specified position specifying histogram has been completed. Notify the unit 260. By this notification, the next position specifying divided region is designated from the position specifying control unit 260, and the histogram generating unit 230 repeats generation of a position specifying histogram corresponding to the designated position specifying divided region. Note that the processing method for generating the position specifying histogram in the histogram generation unit 230 is the same as the processing method for generating the histogram of the entire image in the scene recognition processing, except that the size of the area for generating the histogram is different. .

  The histogram storage unit 280 is a memory such as a DRAM that temporarily stores a histogram of the entire image generated by the histogram generation unit 230 in accordance with control from the histogram generation unit 230. The histogram of the entire image stored in the histogram storage unit 280 is output to the SVM calculation unit 240 in accordance with control from the SVM calculation unit 240.

  Similar to the SVM calculation unit 140 provided in the object position specifying system 10 of the first embodiment, the SVM calculation unit 240 is a histogram generation unit 230 in the scene recognition process according to the control from the position specifying control unit 260. The SVM calculation is performed to compare the histogram of the entire image input from the image and the histogram of each teacher data included in the teacher data group 150, and the similarity is calculated for each category of the object classified in the teacher data group 150. To do. Then, when the SVM calculation unit 240 completes the SVM calculation for the entire histogram of the input image, the object position specifying system 20 displays information representing the similarity for each category of each target calculated by the SVM calculation. It outputs as information determined by performing scene recognition processing. In addition, the SVM calculation unit 240 notifies the position specification control unit 260 that the SVM calculation for performing the scene recognition process is completed. Note that the SVM calculation method in the SVM calculation unit 240 is the same as the SVM calculation unit 140 provided in the object position specifying system 10 of the first embodiment. Since it is the same as the method of, detailed explanation is omitted.

  Further, the SVM calculation unit 240 is different from the SVM calculation unit 140 provided in the object position specifying system 10 of the first embodiment in the process of specifying the position of the object according to the control from the position specifying control unit 260. SVM calculation (hereinafter referred to as “simple SVM calculation”) for comparing each position specifying histogram input from the histogram generation unit 230 and the histogram of the entire image stored in the histogram storage unit 280 is performed, and each position is determined. The similarity with the histogram of the entire image is calculated for each specific divided region. In addition, the SVM calculating unit 240 notifies the position specifying control unit 260 that the simple SVM calculation is completed for each position specifying divided region where the process of specifying the position of the object is performed. By this notification, the next position specifying divided region is designated by the position specifying control unit 260, and the SVM calculating unit 240 repeats the simple SVM calculation for the position specifying histogram of the specified position specifying divided region. Note that the simple SVM calculation method for the position specifying histogram in the SVM calculation unit 240 is also an image in which the histogram of each teacher data included in the teacher data group 150 that performs the SVM calculation process is stored in the histogram storage unit 280. The method is the same as the SVM calculation method in the scene recognition process except that the entire histogram is used.

  In addition, the SVM calculation unit 240 performs processing for specifying the position of the object according to the control from the position specification control unit 260, and after the simple SVM calculation for the position specification histograms of all the position specification divided regions is completed, An SVM operation is performed to compare the position specifying histogram having the highest similarity with the histogram and the histogram of each teacher data included in the teacher data group 150, and the category of the object classified in the teacher data group 150 is calculated. Calculate similarity. Then, when the SVM calculation is completed with respect to the position specifying histogram having the highest similarity to the histogram of the entire image in the simple SVM calculation, the SVM calculating unit 240 calculates the position specifying histogram of the position specifying divided region calculated by the SVM calculation. Information indicating the similarity to the category of the object determined by the scene recognition process is output as a result of the object position specifying system 20 performing the object position specifying process. Note that the SVM calculation method for the position specification histogram in the SVM calculation unit 240 is the same as the SVM calculation unit 140 included in the object position specifying system 10 of the first embodiment. The method is the same as the SVM calculation method in the scene recognition process except that the histogram is used.

  The position specifying control unit 260 is the entire object position specifying system 20, that is, the local feature vector generating unit 110, the quantized vector generating unit 120, the histogram generating unit 230, and the SVM calculating unit included in the object position specifying system 20. Each operation of 240 is controlled. The position specifying control unit 260 includes a histogram generation divided region specifying unit 161 and a position specifying SVM calculation determining unit 262.

  The position specifying control unit 260 has a configuration in which the position specifying control unit 160 provided in the object position specifying system 10 of the first embodiment shown in FIG. 1 is further provided with a position specifying SVM calculation determining unit 262. is there. Note that the histogram generation divided region specifying unit 161 provided in the position specifying control unit 260 is a histogram generation divided region in the position specifying control unit 160 provided in the object position specifying system 10 of the first embodiment shown in FIG. The same operation as the designation unit 161 is performed. Therefore, a detailed description of the operation of the histogram generation divided area designating unit 161 is omitted.

  The position specifying SVM calculation determining unit 262 is used when the SVM calculating unit 240 performs the SVM calculation for each position specifying histogram input from the histogram generating unit 230 in the process of specifying the position of the target in the target position specifying system 20. Is switched to either the histogram of the entire image stored in the histogram storage unit 280 or the histogram of each teacher data included in the teacher data group 150. More specifically, the position specifying SVM calculation determination unit 262 compares the position specifying histogram with each position specifying histogram when the SVM calculating unit 240 performs a simple SVM calculation for each position specifying histogram input from the histogram generation unit 230. The histogram to be switched is switched to the histogram of the entire image stored in the histogram storage unit 280. Further, the position specifying SVM calculation determining unit 262 is similar to the histogram of the entire image after the SVM calculating unit 240 completes the simple SVM calculation for the position specifying histograms of all the position specifying divided regions. When performing the SVM calculation for the position-specific histogram with the highest degree, the histogram that compares the histogram with the position-most histogram with the highest degree of similarity with the histogram of the entire image was determined in the scene recognition process, and the degree of similarity was the highest. Switch to the histogram of the teacher data for each category of the object.

  With such a configuration, in the object position specifying system 20, as in the object position specifying system 10 of the first embodiment, each scene recognition divided region generated by the quantized vector generation unit 120 in the scene recognition process. The position specifying histogram for each position specifying divided region when performing the process of specifying the position of the object is generated using the value of the quantization vector. Further, in the object position specifying system 20, the histogram of the entire image generated by the histogram generating unit 230 is stored in the histogram storing unit 280, and the position of the object is specified using the histogram of the entire image stored in the histogram storing unit 280. Perform the process. More specifically, in the process of scene recognition that is compared in the SVM calculation for the position specifying histogram for each position specifying divided area generated by the histogram generation unit 230 in the process of specifying the position of the object by the object position specifying system 20. Instead of the histograms of the respective teacher data of the category of the object with the highest similarity determined (for example, 1500 histograms included in one category), the histogram of the entire image stored in the histogram storage unit 280 is used. Use. That is, in the object position specifying process by the object position specifying system 20, one histogram representing the entire image generated in the scene recognition process is used instead of a large amount of teacher data used when performing the SVM calculation. Thus, it is possible to easily specify the position of the position specifying divided region where the object determined in the scene recognition process is shown. The reason that one histogram representing the entire image can be used instead of a large amount of teacher data is that the object once determined in the scene recognition processing is reflected in one of the position-specific divided regions. Because it is. Thereby, in the object position specifying system 20, the SVM calculation unit 240 performs the position specifying divided regions generated by the histogram generating unit 230 in order to specify the position specifying histogram having the highest similarity with the histogram of the entire image. It is possible to reduce the time required for the SVM calculation for the position specifying histogram.

  Next, the operation of the object position specifying system 20 will be described. FIG. 7 is a flowchart showing a processing procedure in the object position specifying system 20 according to the second embodiment. Moreover, FIG. 8 is a figure explaining an example of the process which pinpoints the position of a target object in the target object location specifying system 20 of the 2nd embodiment. In the description of the flowchart of the process in the object position specifying system 20 shown in FIG. 7, an example of each process in the object position specifying system 10 of the first embodiment shown in FIGS. An example of processing for specifying the position of the object in the object position specifying system 20 shown in FIG. 8 will be referred to. In the processing in the object position specifying system 20, as in the object position specifying system 10 of the first embodiment, when the object shown in the image is a “dog”, the entire image area is determined. An example will be described in which a position-specific divided area in which the object “dog” is reflected is specified by dividing into nine position-specific divided areas.

  Note that the processing of the object position specifying system 20 includes the same processing as the processing of the object position specifying system 10 of the first embodiment. For this reason, the same processing as the processing of the object position specifying system 10 of the first embodiment is performed in the flowchart showing the processing procedure in the object position specifying system 20 of the second embodiment shown in FIG. The same step number as the step number given to the flowchart showing the processing procedure in the object position specifying system 10 of the first embodiment shown in FIG. 2 is given to the procedure. Therefore, in the description of the flowchart of the process in the object position specifying system 20 shown in FIG.

  When an image is input to the object position specifying system 20, the position specifying control unit 260 is first input in the same manner as the position specifying control unit 160 included in the object position specifying system 10 of the first embodiment. The operation of each component included in the object position specifying system 20 is controlled so that the scene recognition process is performed on the image, and then the object position specifying process is performed (see FIG. 3).

  In the process of scene recognition in the object position specifying system 20, first, in steps S <b> 100 to S <b> 115, the position specifying control unit 260 inputs the image (see FIG. 3A) input to the local feature vector generation unit 110. A local feature vector for each scene recognition divided region is generated, and a quantization vector generation unit 120 generates a quantization vector for each scene recognition divided region, and the value of the generated quantization vector is set for each scene. The quantization vector storage unit 170 stores the recognition divided region.

  Subsequently, in step S120, the position specifying control unit 260 causes the histogram generating unit 230 to calculate the object position specifying system 20 from the quantization vector value for each scene recognition divided region generated by the quantization vector generating unit 120. A histogram representing the entire image (see FIG. 3A) input to is generated. In step S125, the position specifying control unit 260 causes the histogram generation unit 230 to store a histogram representing the entire generated image (see FIG. 3A) in the histogram storage unit 280.

  Subsequently, in step S130, the position specifying control unit 260 causes the SVM calculation unit 240 to display the histogram of the entire image (see FIG. 3A) generated by the histogram generation unit 230 and the teacher data group 150. The SVM calculation for calculating the similarity with the histogram of the teacher data is executed. Thereby, the object position specifying system 20 can determine that “dog” is reflected in the input image (see FIG. 3A), and represents the similarity for each object category. Information is output (see FIG. 4).

  When the scene recognition process from step S100 to step S130 is completed, the object position specifying system 20 determines the position where the determined object is reflected in the image subjected to the scene recognition process from step S300. The position specifying process to be specified is started. In the object position specifying process in the object position specifying system 20, as in the object position specifying system 10 of the first embodiment, first, the quantization vector generation unit 120 stores the quantized vectors in the scene recognition process. A position specifying histogram is generated using the quantization vector value of each scene recognition divided region stored in the unit 170. Thereafter, in the process of specifying the position of the object in the object position specifying system 20, a simple SVM calculation is performed on the generated position specifying histogram, and the entire image stored in the histogram storing unit 280 by the histogram generating unit 230 in the scene recognition process. This is performed for each position-specific divided region using the histogram to be represented. And in the process of specifying the position of the object in the object position specifying system 20, finally, the position specifying divided region where the object determined in the scene recognition process is simply determined by simple SVM calculation is shown. SVM calculation is performed on the position specifying histogram.

  First, in step S300, the position specifying control unit 260 divides the entire region of the input image into the histogram generation unit 230 in the same manner as in step S200 in the object position specifying system 10 of the first embodiment. The first position-specific divided area is designated from among the position-specific divided areas, and the quantization vector value corresponding to the designated first position-specific divided area is acquired from the quantized vector storage unit 170.

  Subsequently, in step S310, the position specifying control unit 260 sends the acquired first position specifying divided region to the histogram generating unit 230 in the same manner as in step S210 in the object position specifying system 10 of the first embodiment. Based on the value of the corresponding quantization vector, a position specifying histogram representing the first position specifying divided region is generated.

  Subsequently, in step S320, the position specifying control unit 260 causes the SVM calculating unit 240 to store the position specifying histogram representing the first position specifying divided region generated by the histogram generating unit 230 and the image stored in the histogram storing unit 280. The simple SVM calculation for calculating the degree of similarity with the histogram representing the whole is executed. As a result, the object position specifying system 20 includes an object having the highest similarity in the scene recognition process (“dog” in the example of the process shown in FIG. 3) in the first position specifying divided region. It is possible to easily determine whether or not the image is shown.

  Subsequently, in step S330, the position specifying control unit 260 determines whether or not the simple determination for all the position specifying divided areas obtained by dividing the entire area of the input image has been completed. If the result of determination in step S330 is that simple determination for all divided position-specific divided areas has not been completed, the process returns to step S300 to specify the next position-specific divided area and all divided positions. The simple discrimination process from step S300 to step S320 is repeated until the simple discrimination for the specific divided region is completed. If the result of determination in step S330 is that simple determination for all divided position specifying divided areas has been completed, simple determination processing in the object position specifying system 20 ends, and the process proceeds to step S340.

  Here, a simple determination process from step S300 to step S320 performed by the object position specifying system 20 will be described. FIG. 8 is a diagram for explaining the concept of processing for simply specifying the position of an object in the object position specifying system 20 according to the second embodiment.

  In the process of specifying the position of the object in the object position specifying system 20, first, the position specifying control unit 260 in steps S300 and S310, steps S200 and S210 in the object position specifying system 10 of the first embodiment. Similarly, the first position-specific divided area A1 is designated among the nine position-specific divided areas (see FIG. 3B) obtained by dividing the entire area of the input image (see FIG. 3A). To do. As a result, the histogram generation unit 230 acquires the quantization vector value corresponding to the first position specifying divided region A1 designated by the position specifying control unit 260 from the quantized vector storage unit 170 (FIG. 5A -1)), a position specifying histogram representing the first position specifying divided area A1 is generated from the obtained quantization vector value corresponding to the first position specifying divided area A1 (FIG. 5 (a-)). 2)). Then, the histogram generation unit 230 outputs a position specifying histogram representing the generated first position specifying divided region A1 to the SVM calculation unit 240.

  Thereafter, in step S <b> 320, the SVM calculation unit 240 acquires a histogram representing the entire image stored in the histogram storage unit 280 in accordance with control from the position specifying control unit 260. Then, the SVM calculation unit 240 performs a simple SVM calculation for calculating the similarity between the histogram representing the entire acquired image and the position specifying histogram representing the first position specifying divided region A1 generated by the histogram generating unit 230. Do. Then, the SVM calculating unit 240 obtains information indicating the similarity to the category “dog” obtained based on the result of the simple SVM calculation calculated for the position specifying divided area A1 in the position specifying divided area A1. Is output as information for easily discriminating whether or not “Dog” is shown in FIG.

  Note that, in the simple SVM calculation by the SVM calculation unit 240 as well, as in the SVM calculation by the SVM calculation unit 240 in the scene recognition process, the position specifying histogram representing the first position specifying divided region A1 generated by the histogram generation unit 230. And the histogram representing the entire acquired image are normalized so that the size of each region is equal, and then the absolute value of the frequency difference between the same classes in each histogram is calculated, and each class is calculated. The difference absolute value of is added. As a result, the SVM calculation unit 240 can easily determine whether or not the “dog” having the highest similarity in the scene recognition process is captured in the first position-specific divided area A1. Information indicating the similarity between the histogram representing the entire image and the position specifying histogram representing the first position specifying divided area A1 is output. In the simple SVM calculation by the SVM calculation unit 240, the position-specific divided area having the smallest value of the calculated difference absolute value is the position-specific divided area in which the object “dog” is reflected. It is possible to determine and output information for specifying the position of the position specifying divided area.

  In addition, when the simple SVM calculation for the position specifying histogram representing the first position specifying divided area A1 designated by the position specifying control section 260 is completed, the SVM calculating section 240 determines the first position specifying divided area A1. The position specifying control unit 260 is notified that the simple SVM calculation for the position specifying histogram is completed. In response to this notification, the position specifying control unit 260 performs the determination in step S330, returns to step S300, and designates the second position specifying divided region A2.

  Then, the histogram generation unit 230 acquires the quantization vector value corresponding to the second position specifying divided region A2 designated by the position specifying control unit 260 from the quantized vector storage unit 170 (FIG. 5 (b-)). 1)), a position specifying histogram representing the second position specifying divided area A2 is generated from the obtained quantization vector value corresponding to the second position specifying divided area A2 (FIG. 5B-2). )reference). Then, the histogram generation unit 230 outputs a position specifying histogram representing the generated second position specifying divided region A2 to the SVM calculation unit 240.

  After that, in step S320, the SVM calculation unit 240, according to the control from the position specifying control unit 260, a histogram representing the entire acquired image and the second position specifying divided region A2 generated by the histogram generating unit 230. A simple SVM calculation is performed to calculate the degree of similarity with the position specifying histogram representing. Then, the SVM calculation unit 240 obtains information indicating the similarity to the category “dog” obtained based on the result of the simple SVM calculation calculated for the position specifying divided area A2 in the position specifying divided area A2. Is output as information for easily discriminating whether or not “Dog” is shown in FIG.

  Thereafter, similarly, the position specifying control unit 260 sequentially designates nine position specifying divided areas (see FIG. 3B) obtained by dividing the entire area of the input image (see FIG. 3A), and the histogram The generation unit 230 sequentially generates a position specifying histogram representing each position specifying divided region designated by the position specifying control unit 260 and outputs the position specifying histogram to the SVM calculation unit 240. Similarly, the SVM calculation unit 240 calculates the degree of similarity between the histogram representing the entire image acquired from the histogram storage unit 280 and the position identification histogram representing each position identification divided region generated by the histogram generation unit 230. Simple SVM calculation is performed. Then, the SVM calculation unit 240 simply determines whether or not “dog” is reflected in each position-specific divided area obtained based on the result of the simple SVM calculation calculated for each position-specific divided area. Output the information to be determined. In FIG. 8, the SVM calculation unit 240 executes a simple SVM calculation that compares the histogram of the entire image stored in the histogram storage unit 280 with each of the position specifying histograms generated by the histogram generation unit 230. An example of a state is shown.

  In addition, the SVM calculation unit 240 selects the position-specific divided area having the highest similarity as the object “dog” based on the information indicating the similarity to the category “dog” in each position-specific divided area. Is identified as a position-specific divided area, and information for specifying the position-specific divided area is output.

  The process so far is the simple determination process of the position specifying divided region in which the object determined in the scene recognition process is included in the object position specifying process by the object position specifying system 20.

  Subsequently, when the simple determination for all the position-specific divided regions is completed in step S330, the position specifying control unit 260 instructs the SVM calculating unit 240 to display the object having the highest similarity in the scene recognition process (see FIG. In the example of the process shown in FIG. 3, the SVM calculation is performed on the position-specific divided area that is simply determined as “dog”). More specifically, in step S340, the position specifying control unit 260 simply determines that the “dog” having the highest similarity in the scene recognition process is captured in the SVM calculating unit 240. SVM calculation for calculating a similarity between a position specifying histogram representing a region and a histogram of each teacher data of the category of the object having the highest similarity, which is included in the teacher data group 150 and is determined in the scene recognition process Is executed. Then, the object position specifying system 20 outputs information representing the similarity with the category “dog” obtained based on the result of the SVM calculation, and the processing in the object position specifying system 20 is completed.

  Note that, in the SVM calculation in step S330 by the SVM calculation unit 240, as well as the SVM calculation by the SVM calculation unit 240 in the scene recognition process, the position specifying histogram representing the position-specific divided areas that are simply determined and the respective teacher data The difference absolute value of the frequency between the same classes in the histogram is calculated, and the difference absolute value of each class is added.

  In this manner, the object position specifying system 20 repeats simple discrimination for all position specifying divided areas obtained by dividing the entire area of the input image (FIG. 3A), thereby specifying each position. Among the divided areas, the position-specific divided area having the highest similarity to the histogram of the entire image is simply identified as the position-specific divided area in which the “dog” determined by the scene recognition process is reflected as the target object. be able to. Then, the object position specifying system 20 uses the position of the position specifying divided area simply specified as the position of the position specifying divided area in which the object determined in the scene recognition process is reflected, and represents information indicating similarity. Can be output.

  According to the present embodiment, the object position specifying system 20 further includes a histogram storage unit (histogram storage unit 280) that stores a histogram representing the entire image generated by the histogram generation unit (histogram generation unit 230). In the position specifying process, the position specifying control unit (position specifying control unit 260) stores each of the histograms representing the position specifying divided regions and the histogram storing unit 280 in the SVM calculating unit (SVM calculating unit 240). An object position specifying system (object position specifying system 20) is configured to execute an SVM calculation that compares a histogram representing the entire image.

  As described above, in the object position specifying system 20 of the second embodiment, the quantized vector generation unit 120 is used in the scene recognition process, as in the object position specifying system 10 of the first embodiment. The generated quantization vector value of each scene recognition divided region is stored in the quantization vector storage unit 170. As a result, in the object position specifying process in the object position specifying system 20 of the second embodiment, the quantized vector storage unit 170 stores the object as in the object position specifying system 10 of the first embodiment. Using the quantization vector value of each scene recognition divided area, a position specifying histogram for each position specifying divided area can be generated with a small amount of processing.

  Further, in the object position specifying system 20 of the second embodiment, the histogram of the entire image generated by the histogram generation unit 230 in the scene recognition process is stored in the histogram storage unit 280. In the object position specifying process in the object position specifying system 20 of the second embodiment, the histogram of each teacher data of the category of the object having the highest similarity determined in the scene recognition process. Instead, simple SVM calculation is performed using the histogram of the entire image stored in the histogram storage unit 280. Thereby, in the object position specifying system 20 of the second embodiment, the position specifying for each position specifying divided area generated by the histogram generation unit 230 is performed to specify the position in the image in which the object is shown. The SVM calculation for the histogram can be performed simply. That is, in the object position specifying process in the object position specifying system 20 of the second embodiment, the SVM calculation for one position specifying histogram is performed in a large amount included in the object category having the highest similarity. It is possible to use only one histogram stored in the histogram storage unit 280 instead of using the teacher data. As a result, in the object position specifying system 20 of the second embodiment, it is possible to narrow down the position specifying divided area where the detailed SVM calculation needs to be performed, and specify the position in the image in which the object is shown. The calculation time required to do this can be further shortened compared to the object position specifying system 10 of the first embodiment.

  Note that in the object position specifying system 20 of the second embodiment as well, as in the object position specifying system 10 of the first embodiment, the SVM calculation unit 240 is information indicating the similarity for each category of the object. In addition, the configuration for outputting the information for specifying the position of the position specifying divided region where the determined object is shown has been described. However, as with the object position specifying system 10 of the first embodiment, other than the SVM calculation unit 240 It is also possible to adopt a configuration in which these components are output. For example, the position specifying SVM calculation determining unit 262 provided in the position specifying control unit 260 is connected to the category of the object obtained from the result of the simple SVM calculation for each position specifying divided region performed by the SVM calculating unit 240. Based on the information indicating the degree of similarity, it may be configured to output information for specifying the position of the position-specific divided region in which the determined object is shown.

  In addition, in the target object specifying system 20 of the second embodiment, the configuration for storing the histogram of the entire image, that is, the histogram storage unit 280 is provided, so that the SVM calculation performed by the SVM calculation unit 240 is simplified. The configuration for shortening the calculation time required for specifying the position in the image in which the object is shown has been described. However, even when the configuration for storing the histogram of the entire image is not provided, the SVM calculation performed by the SVM calculation unit 240 in order to specify the position in the image in which the object is shown can be simplified.

<Third Embodiment>
Next, a third embodiment of the present invention will be described. FIG. 9 is a block diagram showing a schematic configuration of an object position specifying system according to the third embodiment. 9, the object position specifying system 30 includes a local feature vector generating unit 110, a quantized vector generating unit 120, a histogram generating unit 130, an SVM calculating unit 140, a teacher data group 150, and a position specifying control unit. 360, a quantized vector storage unit 170, and a teacher data switching unit 390.

  In addition, in the object position specifying system 30 shown in FIG. 9, the position specifying control unit 160 provided in the object position specifying system 10 of the first embodiment shown in FIG. The teacher data switching unit 390 is provided. Further, the other components provided in the object position specifying system 30 are the same as the components provided in the object position specifying system 10 of the first embodiment shown in FIG. Therefore, in the description of the object position specifying system 30 according to the third embodiment, only components and operations different from those provided in the object position specifying system 10 according to the first embodiment will be described. A detailed description of the same components and operations as those of the object position specifying system 10 of the embodiment will be omitted.

  Similar to the object position specifying system 10 of the first embodiment, the object position specifying system 30 uses a subject (object) shown in the image and a scene in which the image is captured with respect to the input image. Recognizing scene recognition processing is performed, and information on the degree of similarity with each teacher data classified for each type of object is output as information determined by scene recognition processing. In addition, the object position specifying system 30 is a position for specifying the position where the determined object is shown in the image subjected to the scene recognition process, similarly to the object position specifying system 10 of the first embodiment. A specific process is performed, and information indicating the position where the specified object is shown is output.

  The local feature vector generation unit 110 generates a local feature vector of an image input to the object position specifying system 30 for each scene recognition divided region in accordance with the control from the position specifying control unit 360, and generates each of the generated scenes. The value of the local feature vector of the recognition divided region is output to the quantized vector generation unit 120.

  The quantization vector generation unit 120 quantizes the local feature vector value of each scene recognition division region input from the local feature vector generation unit 110 in accordance with the control from the position specification control unit 360. Each quantization vector is generated, and the generated quantization vector value of each scene recognition divided region is output to the histogram generation unit 130 and stored in the quantization vector storage unit 170.

  The histogram generation unit 130 is an image based on the value of the quantization vector for each scene recognition divided region input from the quantization vector generation unit 120 in the process of scene recognition according to the control from the position specifying control unit 360. A histogram representing the entire image is generated, and the generated histogram of the entire image is output to the SVM calculation unit 240. Further, the histogram generation unit 130 performs quantization vector for each scene recognition divided region stored in the quantization vector storage unit 170 in the process of specifying the position of the object according to the control from the position specification control unit 360. A position specifying histogram for each position specifying divided area based on the value of the position specifying divided area is generated, and each of the generated position specifying histograms for each position specifying divided area is output to the SVM calculating unit 240.

  In the scene recognition process according to the control from the position specifying control unit 360, the SVM calculation unit 140 includes the histogram of the entire image input from the histogram generation unit 130 and the histogram of each teacher data included in the teacher data group 150. SVM calculation is performed, and similarity is calculated for each category of objects classified in the teacher data group 150. However, in the object position specifying system 30, each teacher data that the SVM calculation unit 140 compares with the histogram of the entire image is input via the teacher data switching unit 390.

  Further, the SVM calculation unit 140 is included in each position specifying histogram input from the histogram generation unit 130 and the teacher data group 150 in the process of specifying the position of the object according to the control from the position specifying control unit 360. An SVM operation for comparing the histograms of the respective teacher data is performed, and the similarity with the category of the object classified in the teacher data group 150 is calculated for each position specifying divided region. However, in the object position specifying system 30, each teacher data that the SVM calculating unit 140 compares with each position specifying histogram is also input via the teacher data switching unit 390. In the object position specifying system 30, as in the object position specifying system 20 of the second embodiment, first, the SVM calculation for each position specifying histogram is simply performed, and the simple SVM for all the position specifying histograms. After the calculation is completed, the SVM calculation is further performed on the position specifying histogram having the highest similarity, and the similarity with the category of the object classified in the teacher data group 150 is calculated. In the following description, the simple SVM calculation in the object position specifying system 30 is also referred to as “simple SVM calculation”.

  The teacher data switching unit 390 switches each teacher data input to the SVM calculation unit 140 in accordance with the control from the position specifying control unit 360. More specifically, when the SVM calculation unit 140 performs a simple SVM calculation for each position specifying histogram input from the histogram generation unit 130, a histogram of teacher data to be input to the SVM calculation unit 140 is set to a predetermined condition. Only the histogram of the teacher data selected according to Further, after the SVM calculation unit 140 completes the simple SVM calculation for the position specifying histograms of all the position specifying divided areas, the histogram of the teacher data input to the SVM calculating unit 140 when performing the SVM calculation is used for scene recognition. The histogram of all the teacher data of the category of the object with the highest similarity determined in the processing is used.

  Note that the histogram of the teacher data input when the SVM calculation unit 140 performs the simple SVM calculation is a histogram of the teacher data that represents the category of the target object. For example, an image showing the front of the target object, It is the histogram of the teacher data selected by the predetermined conditions which can discriminate | determine easily the target object of the same category by SVM calculation, such as the image in which the side surface was reflected. More specifically, in the case where there are 1500 histograms as one category of teacher data included in the teacher data group 150, for example, 10 histograms of the 1500 histograms are selected according to the conditions described above. select. In the following description, the teacher data representing the category of the selected object is referred to as “extracted teacher data”.

  The position specifying control unit 360 is the entire object position specifying system 30, that is, the local feature vector generating unit 110, the quantized vector generating unit 120, the histogram generating unit 130, and the SVM calculating unit 140 included in the object position specifying system 30. And the operation of the teacher data switching unit 390 are controlled. The position specifying control unit 360 includes a histogram generation divided region specifying unit 161 and a position specifying SVM calculation determining unit 362.

  The position specifying control unit 360 is configured so that the position specifying SVM calculation determining unit 262 in the position specifying control unit 260 included in the object position specifying system 10 of the first embodiment shown in FIG. This is a configuration that replaces the portion 362. Note that the histogram generation divided region specifying unit 161 included in the position specifying control unit 360 includes a histogram generation divided region specifying unit 161 and a first one in the position specifying control unit 160 included in the object position specifying system 10 according to the first embodiment. The same operation as that of the histogram generation divided region specifying unit 161 in the position specifying control unit 260 provided in the object position specifying system 20 of the second embodiment is performed. Therefore, a detailed description of the operation of the histogram generation divided area designating unit 161 is omitted.

  The position specifying SVM calculation determining unit 362 is used when the SVM calculating unit 140 performs the SVM calculation for each position specifying histogram input from the histogram generating unit 130 in the process of specifying the position of the target in the target position specifying system 30. Is selected from either the histogram of each teacher data included in the teacher data group 150 or the histogram of extracted teacher data selected according to a predetermined condition. Accordingly, the teacher data switching unit 390 performs a histogram of all the teacher data of objects of the same category included in the teacher data group 150 in the object position specifying process in the object position specifying system 30 or a predetermined value. One of the histograms of a part of the teacher data in the teacher data group 150 selected according to the condition is output to the SVM calculation unit 140. More specifically, the position specifying SVM calculation determining unit 362 compares the position specifying SVM calculation unit 140 with each position specifying histogram when the SVM calculating unit 140 performs a simple SVM calculation for each position specifying histogram input from the histogram generating unit 130. The teacher data switching unit 390 is controlled so that the histogram to be used is a histogram of a part of the teacher data representing the category of the object having the highest similarity determined in the scene recognition process. In addition, the position specifying SVM calculation determining unit 362, when the SVM calculating unit 140 further performs the SVM calculation after the SVM calculating unit 140 completes the simple SVM calculation for the position specifying histograms of all the position specifying divided regions. Histograms of all the teacher data of the category of the object having the highest similarity determined in the scene recognition process are compared with the position identification histogram having the highest similarity with the teacher data compared in the simple SVM calculation. In this manner, the teacher data switching unit 390 is controlled.

  With such a configuration, in the object position specifying system 30, as in the object position specifying system 10 of the first embodiment, each scene recognition divided region generated by the quantized vector generation unit 120 in the scene recognition process. The position specifying histogram for each position specifying divided region when performing the process of specifying the position of the object is generated using the value of the quantization vector. Further, the object position specifying system 30 selects either one of the extracted teacher data or all the teacher data for the teacher data histogram used when performing the SVM calculation in the object position specifying process. Using the teacher datum, the process of specifying the position of the object is performed. More specifically, in the process of identifying a target by the target object specifying system 30, in the process of scene recognition to be compared in the SVM calculation for the position specifying histogram for each position specifying divided region generated by the histogram generation unit 130. Extracted teacher data selected according to a predetermined condition, instead of the histograms of all the teacher data of the category of the object with the highest similarity determined (for example, 1500 histograms included in one category) The histogram is used. That is, in the object position specifying process by the object position specifying system 30, instead of a large amount of teacher data used when performing the SVM calculation, the object having the highest similarity determined in the scene recognition process is used. Using the histogram of extracted teacher data representing the category, the position of the position-specific divided area where the object determined in the scene recognition process is reflected can be easily specified. Thereby, in the object position specifying system 30, the histogram generating unit 130, which is performed by the SVM calculating unit 140 in order to specify the position specifying histogram having the highest similarity with the histogram of the extracted teacher data representing the category of the object, is provided. It is possible to reduce the time required for the SVM calculation for the generated position specifying histogram for each position specifying divided region.

  Next, the operation of the object position specifying system 30 will be described. FIG. 10 is a flowchart showing a processing procedure in the object position specifying system 30 of the third embodiment. FIG. 11 is a diagram illustrating an example of processing for specifying the position of an object in the object position specifying system 30 according to the third embodiment. In the description of the flowchart of the process in the object position specifying system 30 shown in FIG. 10, an example of each process in the object position specifying system 10 of the first embodiment shown in FIGS. An example of processing for specifying the position of an object in the object position specifying system 30 shown in FIG. 11 will be referred to. And also in the processing in the object position specifying system 30, as in the object position specifying system 10 of the first embodiment, when the object shown in the image is “dog”, the entire image area is determined. An example will be described in which a position-specific divided area in which the object “dog” is reflected is specified by dividing into nine position-specific divided areas.

  Note that the processing of the object position specifying system 30 includes the same processing as the processing of the object position specifying system 10 of the first embodiment. For this reason, the same processing as the processing of the object position specifying system 10 of the first embodiment is performed in the flowchart showing the processing procedure in the object position specifying system 30 of the third embodiment shown in FIG. The same step number as the step number given to the flowchart showing the processing procedure in the object position specifying system 10 of the first embodiment shown in FIG. 2 is given to the procedure. Therefore, in the description of the flowchart of the process in the object position specifying system 30 illustrated in FIG. 10, the detailed description regarding the procedure for performing the same process as the process of the object position specifying system 10 of the first embodiment is omitted.

  When an image is input to the object position specifying system 30, the position specifying control unit 360 is first input in the same manner as the position specifying control unit 160 provided in the object position specifying system 10 of the first embodiment. The operation of each component included in the object position specifying system 30 is controlled so that the scene recognition process is performed on the image, and then the object position specifying process is performed (see FIG. 3).

  In the process of scene recognition in the object position specifying system 30, first, in steps S100 to S115, the position specifying control unit 360 inputs the image (see FIG. 3A) input to the local feature vector generating unit 110. A local feature vector for each scene recognition divided region is generated, and a quantization vector generation unit 120 generates a quantization vector for each scene recognition divided region, and the value of the generated quantization vector is set for each scene. The quantization vector storage unit 170 stores the recognition divided region.

  Subsequently, in step S120 to step S130, the position specifying control unit 360 inputs the image (see FIG. 3A) input to the histogram generation unit 130 from the value of the quantization vector for each scene recognition divided region. A histogram representing the whole is generated, and the SVM calculation unit 140 is caused to execute an SVM calculation on the histogram of the entire generated image (see FIG. 3A). Thereby, the object position specifying system 30 can determine that “dog” is reflected in the input image (see FIG. 3A), and represents the similarity of each object for each category. Information is output (see FIG. 4).

  When the scene recognition process from step S100 to step S130 is completed, the object position specifying system 30 determines the position where the determined object is reflected in the image subjected to the scene recognition process from step S400. The position specifying process to be specified is started. In the object position specifying process in the object position specifying system 30, first, scene recognition is performed in the same manner as in the object position specifying system 10 of the first embodiment and the object position specifying system 20 of the second embodiment. In the processing, the quantization vector generation unit 120 generates a position specifying histogram using the quantization vector value of each scene recognition divided region stored in the quantization vector storage unit 170. Thereafter, in the process of specifying the position of the object in the object position specifying system 30, the simple SVM calculation for the generated position specifying histogram is performed for each position specifying divided region using the histogram of the extracted teacher data. Then, in the object position specifying process in the object position specifying system 30, finally, as in the object position specifying system 20 of the second embodiment, the scene recognition of the scene recognition determined simply by the simple SVM calculation is performed. The SVM calculation is performed on the position specifying histogram representing the position specifying divided region in which the object determined in the process is shown.

  First, in step S400 to step S410, the position specifying control unit 360 causes the histogram generating unit 130 to specify the first position as in step S200 to step S210 in the object position specifying system 10 of the first embodiment. Designates a divided area, acquires the value of the quantization vector corresponding to the designated first position-specific divided area from the quantization vector storage unit 170, and generates a position-specific histogram representing the first position-specific divided area Let

  Subsequently, in step S420, the position specifying control unit 360 causes the SVM calculating unit 140 to resemble the position specifying histogram representing the first position specifying divided region generated by the histogram generating unit 130 and the histogram of the extracted teacher data. A simple SVM calculation for calculating the degree is executed. As a result, the object position specifying system 30 includes an object having the highest similarity in the scene recognition process (“dog” in the example of the process shown in FIG. 3) in the first position specifying divided region. It is possible to easily determine whether or not the image is shown.

  Subsequently, in step S430, the position specifying control unit 360 determines whether or not the simple determination for all the position specifying divided areas obtained by dividing the entire area of the input image has been completed. If the result of determination in step S430 is that simple determination has not been completed for all divided position-specific divided areas, the process returns to step S400 to specify the next position-specific divided area and all divided positions. The simple discrimination process from step S400 to step S420 is repeated until the simple discrimination for the specific divided region is completed. As a result of the determination in step S430, when the simple determination for all the divided position specifying divided regions is completed, the simple determination process in the object position specifying system 30 is ended, and the process proceeds to step S440.

  Here, a simple determination process from step S400 to step S420 performed by the object position specifying system 30 will be described. FIG. 11 is a diagram illustrating the concept of processing for simply specifying the position of an object in the object position specifying system 30 according to the third embodiment.

  In the process of specifying the position of the object in the object position specifying system 30, first, the position specifying control unit 360 in steps S400 and S410, steps S200 and S210 in the object position specifying system 10 of the first embodiment. Similarly, the first position-specific divided area A1 is designated among the nine position-specific divided areas (see FIG. 3B) obtained by dividing the entire area of the input image (see FIG. 3A). To do. Thereby, the histogram generation unit 130 acquires the value of the quantization vector corresponding to the first position specifying divided region A1 specified by the position specifying control unit 360 from the quantized vector storage unit 170 (FIG. 5A -1))) A position specifying histogram representing the first position specifying divided area A1 is generated (see FIG. 5A-2), and the position specifying histogram representing the generated first position specifying divided area A1 is generated. Is output to the SVM calculation unit 140.

  After that, in step S420, the SVM calculation unit 140 switches the teacher data switching with the position specifying histogram representing the first position specifying divided area A1 generated by the histogram generating unit 130 in accordance with the control from the position specifying control unit 360. A simple SVM calculation is performed to calculate the degree of similarity with the histogram of each extracted teacher data input via the unit 390. Then, the SVM calculation unit 140 obtains information indicating the similarity with the category “dog” obtained based on the result of the simple SVM calculation calculated for the position specifying divided area A1 in the position specifying divided area A1. Is output as information for easily discriminating whether or not “Dog” is shown in FIG.

  Note that, in the simple SVM calculation performed by the SVM calculation unit 140, as in the SVM calculation performed by the SVM calculation unit 140 in the scene recognition process, the position specifying histogram representing the first position specifying divided region A1 generated by the histogram generation unit 130 is used. And the histograms of the respective extracted teacher data are normalized so that the sizes of the regions are equal to each other, and then the difference absolute values of the frequencies of the same classes in the respective histograms are added. Thereby, the SVM calculation unit 140 can easily determine whether or not the “dog” having the highest similarity in the scene recognition process is captured in the first position-specific divided area A1. Information indicating the similarity between the histogram of each extracted teacher data and the position specifying histogram representing the first position specifying divided area A1 is output. In the simple SVM calculation by the SVM calculation unit 140, the position-specific divided area having the smallest value of the calculated difference absolute value addition value is the position-specific divided area in which the “dog” that is the object is reflected. It discriminate | determines and outputs the information which pinpoints the position of the position specific division area.

  In addition, when the simple SVM calculation for the position specifying histogram representing the first position specifying divided area A1 designated by the position specifying control section 360 is completed, the SVM calculating section 140 determines the first position specifying divided area A1. The position specifying control unit 360 is notified that the simple SVM calculation for the position specifying histogram is completed. In response to this notification, the position specifying control unit 360 performs the determination in step S430, returns to step S400, and specifies the second position specifying divided region A2.

  Then, the histogram generation unit 130 acquires the quantization vector value corresponding to the second position specifying divided region A2 designated by the position specifying control unit 360 from the quantized vector storage unit 170 (FIG. 5B-b). (See 1)) A position specifying histogram representing the second position specifying divided area A2 is generated (see FIG. 5B-2), and a position specifying histogram representing the generated second position specifying divided area A2 is generated. , Output to the SVM calculation unit 140.

  After that, in step S420, the SVM calculation unit 140 switches the position specifying histogram representing the second position specifying divided region A2 generated by the histogram generating unit 130 according to the control from the position specifying control unit 360, and teacher data switching. A simple SVM calculation is performed to calculate the degree of similarity with the histogram of each extracted teacher data input via the unit 390. Then, the SVM calculation unit 140 obtains information indicating the similarity with the category “dog” obtained based on the result of the simple SVM calculation calculated for the position specifying divided area A2 in the position specifying divided area A2. Is output as information for easily discriminating whether or not “Dog” is shown in FIG.

  Thereafter, similarly, the position specifying control unit 360 sequentially designates nine position specifying divided areas (see FIG. 3B) obtained by dividing the entire area of the input image (see FIG. 3A), and the histogram The generating unit 130 sequentially generates a position specifying histogram representing each position specifying divided region designated by the position specifying control unit 360 and outputs the position specifying histogram to the SVM calculating unit 140. Similarly, the SVM calculating unit 140 includes a position specifying histogram representing each position specifying divided region generated by the histogram generating unit 130, and a histogram of each extracted teacher data input via the teacher data switching unit 390. A simple SVM calculation for calculating the degree of similarity is performed. Then, the SVM calculating unit 140 simply determines whether or not “dog” is reflected in each position specifying divided area obtained based on the result of the simple SVM calculation calculated for each position specifying divided area. Output the information to be determined. In FIG. 11, the SVM calculation unit 140 includes each of the histograms of extracted teacher data in which 10 histograms are selected from the 1500 histograms included in the category whose target is “dog”, and the histogram generation unit. An example of a state in which a simple SVM calculation that compares each of the position specifying histograms generated by 130 is executed is shown.

  Further, the SVM calculation unit 140 selects the position-specific divided area having the highest similarity as the object “dog” based on the information indicating the similarity to the category “dog” in each position-specific divided area. Is identified as a position-specific divided area, and information for specifying the position-specific divided area is output.

  The process so far is the simple determination process of the position-specific divided region in which the object determined in the scene recognition process is included in the object position specifying process by the object position specifying system 30.

  Subsequently, when the simple determination for all the position-specific divided regions is completed in step S430, the position-specific control unit 360 instructs the SVM calculation unit 140 to display the object having the highest similarity in the scene recognition process (see FIG. In the example of the process shown in FIG. 3, the SVM calculation is performed on the position-specific divided area that is simply determined as “dog”). More specifically, in step S440, the position specifying control unit 360 simply determines that the “dog” having the highest similarity in the scene recognition process is captured in the SVM calculating unit 140. SVM calculation for calculating a similarity between a position specifying histogram representing a region and a histogram of each teacher data of the category of the object having the highest similarity, which is included in the teacher data group 150 and is determined in the scene recognition process Is executed. Then, the object position specifying system 30 outputs information representing the similarity with the category “dog” obtained based on the result of the SVM calculation, and the processing in the object position specifying system 30 is completed.

  Note that, in the SVM calculation in step S430 by the SVM calculation unit 140 as well, as in the SVM calculation by the SVM calculation unit 140 in the scene recognition process, the position specification histogram representing the position determination divided region that is simply determined and the respective teacher data The difference absolute value of the frequency between the same classes in the histogram is added.

  In this way, the object position specifying system 30 also repeats simple determination for all position specifying divided areas obtained by dividing the entire area of the input image (FIG. 3A), thereby specifying each position. Among the divided areas, the position-specific divided area having the highest similarity to the histogram of the extracted teacher data is simply identified as the position-specific divided area in which the “dog” identified by the scene recognition process is reflected as the target object. can do. Then, in the object position specifying system 30, information indicating similarity is obtained by using the position of the position specifying divided area simply specified as the position of the position specifying divided area where the object determined in the scene recognition processing is reflected. Can be output.

  According to the present embodiment, in the position specifying process, the position specifying control unit (position specifying control unit 360) causes the SVM calculating unit 140 to include each of the histogram representing the position specifying divided region and a plurality of teacher data. An object position specifying system (object position specifying system 30) is configured to execute an SVM operation for comparing each of histograms of a part of teacher data (extracted teacher data) selected according to a predetermined condition. .

  As described above, in the object position specifying system 30 according to the third embodiment, the quantized vector generation unit 120 performs the scene recognition processing in the same manner as the object position specifying system 10 according to the first embodiment. The generated quantization vector value of each scene recognition divided region is stored in the quantization vector storage unit 170. As a result, the object position specifying process in the object position specifying system 30 of the third embodiment is also stored in the quantized vector storage unit 170 as in the object position specifying system 10 of the first embodiment. Using the quantization vector value of each scene recognition divided area, a position specifying histogram for each position specifying divided area can be generated with a small amount of processing.

  Further, in the target object specifying system 30 of the third embodiment, all the teacher data of the category of the target object having the highest similarity determined in the scene recognition process in the target object specifying process is obtained. A simple SVM calculation is performed using a histogram of extracted teacher data representing the category of the object instead of the histogram. Thereby, in the target object specifying system 30 of the third embodiment, the position specifying for each position specifying divided region generated by the histogram generating unit 130 is performed to specify the position in the image in which the target is shown. The SVM calculation for the histogram can be performed simply. That is, in the object position specifying process in the object position specifying system 30 according to the third embodiment, the SVM calculation for one position specifying histogram is performed in a large amount included in the object category having the highest similarity. It is possible to use a histogram of a part of teacher data selected according to a predetermined condition from teacher data included in the same object category. As a result, in the target object specifying system 30 of the third embodiment, it is possible to narrow down the position specifying divided area where the detailed SVM calculation needs to be performed, and to specify the position in the image in which the target is shown. The calculation time required to do this can be further shortened compared to the object position specifying system 10 of the first embodiment.

  Note that in the object position specifying system 30 of the third embodiment as well, as in the object position specifying system 10 of the first embodiment, the SVM calculation unit 140 is information indicating the similarity for each category of the object. In addition, the configuration for outputting the information for specifying the position of the position specifying divided region where the determined object is shown has been described. However, as with the object position specifying system 10 of the first embodiment, other than the SVM calculation unit 140 It is also possible to adopt a configuration in which these components are output. For example, the position specifying SVM calculation determining unit 362 included in the position specifying control unit 360 may determine the category of the object obtained from the result of the simple SVM calculation for each position specifying divided region performed by the SVM calculating unit 140. Based on the information indicating the degree of similarity, it may be configured to output information for specifying the position of the position-specific divided region in which the determined object is shown.

  Note that the object position specifying system 30 according to the third embodiment includes the teacher data switching unit 390, and the position specifying control unit 360 controls the teacher data switching unit 390 so that the object position specifying process 30 is performed. A configuration has been described in which the histogram of teacher data used for the simple SVM calculation performed by the SVM calculation unit 140 on each position specifying histogram is switched to either extracted teacher data or all teacher data. However, the method of switching the histogram of the teacher data used when the SVM calculation unit 140 performs the simple SVM calculation is not limited to the method by the teacher data switching unit 390. For example, information such as a flag indicating whether or not the teacher data representing the category of the object, that is, whether or not the extracted teacher data is selected, is included in each teacher data. Whether the position specifying control unit 360 uses teacher data including a flag indicating that it is selected as extracted teacher data, as the teacher data used when the SVM calculation unit 140 performs the simple SVM calculation, Or it is set as the structure which designates whether the teacher data containing the flag showing not being selected as extraction teacher data are used. Also with this configuration, it is possible to switch the histogram of teacher data used for the simple SVM calculation performed by the SVM calculation unit 140 on each position specifying histogram. With this configuration, the position specification control unit 360 can directly instruct the teacher data used when performing the simple SVM calculation to the SVM calculation unit 140, and the teacher data switching unit 390 is connected to the object position specifying system 30. It does not have to be provided.

  Note that the target object specifying system 20 of the second embodiment includes a histogram storage unit 280 that stores a histogram of the entire image generated by the histogram generation unit 230 in the scene recognition process, and the target determined by the scene recognition process. In order to easily specify the position in the image where the object is shown, the operation using the histogram of the whole stored image was shown. In addition, the object position specifying system 30 according to the third embodiment includes a teacher data switching unit 390 that switches teacher data used in the simple SVM calculation process, and an image in which the object determined by the scene recognition process is shown. In order to easily identify the position in the image, an operation using a histogram of extracted teacher data representing the category of the object having the highest similarity in the scene recognition process is shown. However, the configuration for easily specifying the position in the image in which the object determined by the scene recognition process is reflected is not limited to the configuration of the second embodiment or the third embodiment. Absent. For example, as a configuration for easily specifying a position in an image in which an object determined by scene recognition processing is shown, the configuration of the object position specifying system 20 shown in the second embodiment, The configuration of the object position specifying system 30 shown in the embodiment can be configured to be provided at the same time. In this case, for example, the second embodiment is a method for simply specifying the position in the image in which the target object is captured based on the degree of similarity between the teacher data and the target object determined by the scene recognition process. It is possible to switch to one of the operations of the object position specifying system 20 of the form or the object position specifying system 30 of the third embodiment.

<Fourth Embodiment>
Here, a description will be given of a fourth embodiment of the present invention that includes the configuration of the object position specifying system 20 of the second embodiment and the configuration of the object position specifying system 30 of the third embodiment at the same time. . FIG. 12 is a block diagram showing a schematic configuration of an object position specifying system according to the fourth embodiment. 12, the object position specifying system 40 includes a local feature vector generating unit 110, a quantized vector generating unit 120, a histogram generating unit 230, an SVM calculating unit 140, a teacher data group 150, and a position specifying control unit. 460, a quantized vector storage unit 170, a histogram storage unit 280, and a teacher data switching unit 490.

  The object position specifying system 40 shown in FIG. 12 includes the components of the object position specifying system 20 of the second embodiment shown in FIG. 6 and the object of the third embodiment shown in FIG. This is a configuration combining the components of the position specifying system 30. Therefore, in the description of the object position specifying system 40 according to the fourth embodiment, the components of the object position specifying system 20 according to the second embodiment and the object position specifying system 30 according to the third embodiment. Only components and operations different from the components will be described.

  The teacher data switching unit 490 switches the histogram input to the SVM calculation unit 140 in accordance with control from the position specifying control unit 460. More specifically, the histogram input when the SVM calculation unit 140 performs the simple SVM calculation for each position specifying histogram input from the histogram generation unit 230 is used for each teacher data included in the teacher data group 150. The histogram is one of a histogram, a histogram of the entire image stored in the histogram storage unit 280, or a histogram of extracted teacher data selected according to a predetermined condition.

  The position specifying control unit 460 is the entire object position specifying system 40, that is, the local feature vector generating unit 110, the quantized vector generating unit 120, the histogram generating unit 230, and the SVM calculating unit 140 provided in the object position specifying system 40. And the operation of the teacher data switching unit 490 are controlled. The position specifying control unit 460 includes a histogram generation divided region specifying unit 161 and a position specifying SVM calculation determining unit 462.

  The position specifying SVM calculation determining unit 462 controls the teacher data switching unit 490 in the object position specifying process in the target object specifying system 40, so that the SVM calculating unit 140 is input from the histogram generating unit 230. The histogram for performing the SVM calculation for each position specifying histogram is switched. In addition, the operation | movement at the time of switching the histogram which the position specific SVM calculation determination part 462 inputs into the SVM calculation part 140 is the position specific SVM in the position specific control part 260 with which the target object position specific system 20 of 2nd Embodiment was equipped. Since the operation determination unit 262 and the position specifying SVM calculation determining unit 362 in the position specifying control unit 360 provided in the object position specifying system 30 of the third embodiment can be easily understood, Detailed description is omitted.

  With such a configuration, in the object position specifying system 40, either the object position specifying system 20 of the second embodiment or the object position specifying system 30 of the third embodiment is controlled by the position specifying control unit 460. It is possible to perform the same operation as either one. The operation of the object position specifying system 40 is the same as the operation of the object position specifying system 20 of the second embodiment shown in FIG. 7 or the object position specifying system 30 of the third embodiment shown in FIG. Therefore, detailed description is omitted.

  According to the present embodiment, in the object position specifying system 40, the histogram storage unit 280 that stores the histogram representing the entire image generated by the histogram generation unit 230, and the entire image stored in the histogram storage unit 280 are stored. A teacher data switching unit (teacher data switching unit 490) for selecting and outputting either a histogram to be represented or a histogram of extracted teacher data selected according to a predetermined condition from a plurality of teacher data In the position specifying process, the position specifying control unit (position specifying control unit 460) causes the SVM calculation unit 140 to control each of the histograms indicating the position specifying divided regions and the teacher data switching unit to control the teacher data. Histogram output from the switching unit 490 (in the histogram storage unit 280 Histogram represents the overall presence images, or to perform an SVM operation for comparing the histograms) of the extraction teacher data, is composed of (the object location system 40).

  As described above, in the object position specifying system 40 of the fourth embodiment, each scene stored in the quantized vector storage unit 170 is the same as the object position specifying system 10 of the first embodiment. Using the value of the quantization vector of the recognition divided area, a position specifying histogram for each position specifying divided area can be generated with a small amount of processing.

  Further, in the object position specifying system 40 of the fourth embodiment, the same method as that of either the object position specifying system 20 of the second embodiment or the object position specifying system 30 of the third embodiment. Thus, the SVM calculation for the position specifying histogram for each position specifying divided region generated by the histogram generation unit 230, which is performed to specify the position in the image in which the object is captured, can be easily performed. As a result, the object position specifying system 40 of the fourth embodiment is the same as the object position specifying system 20 of the second embodiment or the object position specifying system 30 of the third embodiment. It is possible to narrow down the position specifying divided area where the SVM calculation needs to be performed, and the calculation time required to specify the position in the image in which the object is shown is determined by the object position specifying system 10 of the first embodiment. Can be further shortened.

  Note that the simple SVM calculation performed to specify the position in the image in which the object is shown is the operation of the object position specifying system 20 of the second embodiment or the object position specifying system 30 of the third embodiment. It is conceivable that which of the above operations is performed is switched according to the degree of similarity between the teacher data and the object determined by the scene recognition process, for example. More specifically, when the similarity with the category of the object determined in the scene recognition process is 80% or more, a simple SVM calculation is performed by the operation of the object position specifying system 20 of the second embodiment, and the scene When the similarity to the category of the object determined in the recognition process is 60% or more and less than 80%, the simple SVM calculation may be performed by the operation of the object position specifying system 30 of the third embodiment. it can. Further, when the similarity with the category of the object determined in the scene recognition process is less than 60%, the simple SVM calculation is not performed, and the normal SVM calculation is performed by the operation of the object position specifying system 10 of the first embodiment. Can be done.

  As described above, according to the mode for carrying out the present invention, the quantization vector for storing the quantization vector value of each scene recognition divided region generated by the quantization vector generation unit in the scene recognition processing. A storage unit is provided. Further, in the embodiment for carrying out the present invention, an image in which the object determined by the scene recognition process is shown using the quantization vector value of each scene recognition divided region stored in the quantization vector storage unit. A position specifying histogram is generated for each position specifying divided area having a predetermined size, which is used to specify the position within the area. Thereby, in the form for implementing this invention, after performing the process of scene recognition with respect to the input image, each position specification performed in order to specify the position in the image in which the target object is reflected is shown. The process of specifying the position of the object with respect to the divided area can be performed with fewer processes than when the process equivalent to the scene recognition process is performed again. Thereby, in the form for implementing this invention, the calculation time required in order to pinpoint the position in the image in which the target object is reflected can be shortened.

  Further, according to the embodiment for carrying out the present invention, the histogram of the whole image generated by the histogram generation unit in the scene recognition process or the category of the object having the highest similarity in the scene recognition process is represented. The SVM calculation is performed using the histogram of the teacher data of the part. Thereby, in the form for implementing this invention, the SVM calculation with respect to the position specific histogram for every position specific division area can be performed simply, and the image which the target object discriminate | determined by the process of scene recognition is reflected The position inside can be easily specified. That is, in the embodiment for carrying out the present invention, in order to specify the position in the image in which the object determined by the scene recognition process is shown, the SVM calculation process performed for each position specifying divided region is performed. Instead of using a histogram of a large amount of teacher data included in the category of the object having the highest similarity in the scene recognition process, it can be easily performed using a small number of histograms. Thereby, in the form for implementing this invention, the calculation time required in order to pinpoint the position in the image in which the target object is reflected can be shortened.

  In the present embodiment, after a series of processing (that is, processing of generating a position specifying histogram and processing of SVM calculation) for one position specifying divided region is completed, in response to a notification indicating that the process is completed, The operation for executing the following series of processes has been described. However, a series of processing operations for each position-specific divided region is not limited to the operations described in the embodiment for carrying out the present invention. For example, after the generation of the position specifying histogram representing the first position specifying divided area is completed, the position specifying representing the second position specifying divided area is performed simultaneously with the SVM calculation for the first position specifying divided area. It can also be controlled to generate a histogram. That is, for example, the SVM calculation for the first position specific divided area and the generation of the histogram for the second position specific divided area may be executed in parallel.

  In the present embodiment, the example in which the input image is divided into 81 scene recognition divided areas and divided into nine position-specific divided areas has been described, but scene recognition for dividing the input image is described. The numbers of the divided areas and the position-specific divided areas are not limited to the numbers described in the embodiment for carrying out the present invention.

  The embodiment of the present invention has been described above with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes various modifications within the scope of the present invention. It is.

10, 20, 30, 40 ... object position specifying system 110 ... local feature vector generation unit 120 ... quantization vector generation unit 130, 230 ... histogram generation unit 140, 240 ... SVM calculation Unit 150 ... Teacher data group 160, 260, 360, 460 ... Position specifying control unit 161 ... Histogram generation divided area specifying unit (position specifying control unit)
170... Quantized vector storage units 262, 362, 462... Location specifying SVM calculation determining unit (position specifying control unit)
280: Histogram storage unit 390, 490 ... Teacher data switching unit

Claims (5)

The entire area of the input image is divided into a plurality of first areas having a predetermined first size, and the image data included in the first area is divided for each of the divided first areas. A local feature vector generating unit for generating a local feature vector representing a local feature in
A quantization vector generation unit that quantizes the value of the local feature vector of each of the first regions generated by the local feature vector generation unit and generates a quantization vector corresponding to each of the first regions; ,
A quantization vector storage unit that stores the value of each of the quantization vectors generated by the quantization vector generation unit for each of the first regions;
A histogram generation unit that generates a histogram representing a whole or a partial region of the image from the value of the quantization vector for each of the first regions;
An SVM calculation unit that performs a support vector machine (SVM) calculation on the histogram generated by the histogram generation unit;
A scene recognition unit that controls each of the local feature vector generation unit, the quantization vector generation unit, the histogram generation unit, and the SVM calculation unit to recognize a scene of the image in which an object is copied. After the process is executed, the object determined in the scene recognition process includes a plurality of second areas obtained by dividing the entire area of the image into a predetermined second size larger than the first area. A position specifying control unit for executing a position specifying process for specifying at which position in the area;
With
The position specifying control unit includes:
In the scene recognition process,
The histogram generation unit generates a histogram representing the entire image from the value of the quantization vector for each of the first regions, and the SVM calculation unit includes a histogram representing the entire image, and a plurality of histograms. The histogram of the image is compared with each of the histograms of the plurality of teacher data that are grouped according to the type of the object,
In the position specifying process,
The histogram generation unit generates a histogram representing each image of the second region from the value of the quantization vector for each of the first regions stored in the quantization vector storage unit, Causing the SVM calculation unit to execute an SVM calculation for each of the histograms representing the second region;
An object location system characterized by the above. A histogram storage unit that stores the histogram generated by the histogram generation unit and representing the entire image;
Further comprising
The position specifying control unit includes:
In the position specifying process,
Causing the SVM calculation unit to execute an SVM calculation that compares each of the histograms representing the second region with a histogram representing the whole of the image stored in the histogram storage unit;
The object position specifying system according to claim 1. The position specifying control unit includes:
In the position specifying process,
The SVM calculating unit compares each of the histograms representing the second region with each of the histograms of a part of the teacher data selected according to a predetermined condition among the plurality of teacher data. To perform the operation,
The object position specifying system according to claim 1 or 2, characterized in that: A histogram storage unit that stores the histogram generated by the histogram generation unit and representing the entire image;
Select either one of the histogram representing the entire image stored in the histogram storage unit, or a part of the teacher data selected from a plurality of the teacher data according to a predetermined condition. An output teacher data switching unit;
Further comprising
The position specifying control unit includes:
In the position specifying process,
Causing the SVM calculation unit to execute an SVM calculation that compares each of the histograms representing the second region with a histogram output from the teacher data switching unit by controlling the teacher data switching unit;
The object position specifying system according to any one of claims 1 to 3, characterized in that The entire area of the input image is divided into a plurality of first areas having a predetermined first size, and the image data included in the first area is divided for each of the divided first areas. A local feature vector generating unit that generates a local feature vector that represents a local feature in, and quantizing the value of the local feature vector of each of the first regions generated by the local feature vector generating unit, A quantization vector generation unit that generates a quantization vector corresponding to the first region, and a quantum that stores the value of each of the quantization vectors generated by the quantization vector generation unit for each first region A quantized vector storage unit, a histogram generating unit that generates a histogram representing the whole or a partial region of the image from the value of the quantization vector for each of the first regions, and the hiss And SVM operation unit for performing support vector machine (SVM) operation on the histogram grams generator has generated, and the local feature vector generating unit, and the quantized vector generation unit, and the histogram generating unit, and the SVM computation unit Each of the images is controlled to execute a scene recognition process for recognizing the scene of the image in which the object is copied, and then the object determined in the scene recognition process defines the entire area of the image. A position specifying control unit for executing a position specifying process for specifying in which position of the plurality of second areas divided into a predetermined second size larger than the first area. In an object location system comprising:
The position specifying control unit
In the scene recognition process,
A step of causing the histogram generation unit to generate a histogram representing the entire image from the value of the quantization vector for each of the first regions; and a histogram representing the entire image to the SVM calculation unit; A procedure for executing an SVM operation for comparing each of a plurality of histograms of a plurality of teacher data in which histograms of a plurality of images are classified and grouped for each type of object;
Including
In the position specifying process,
A step of causing the histogram generation unit to generate a histogram representing an image of each of the second regions from the value of the quantization vector for each of the first regions stored in the quantization vector storage unit; , Causing the SVM calculation unit to execute SVM calculation for each of the histograms representing the second region;
including,
An object position specifying method characterized by the above.

Images