JP2012022419A - Learning data creation device, learning data creation method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To automatically create learning data from image data collected on the Web without human assistance.SOLUTION: A learning data creation device determines a cluster having a small number of area image distributions from the area image classification status for clusters based on the number of area images belonging to each of the clusters, and removes area images belonging to the cluster from the image data, so as to create learning data. Thus, image areas without commonality throughout images are removed, and image areas (objects indicating a keyword) with commonality among the images are left in the learning data created from the image data collected by keyword retrieval.

Description

  The present invention relates to a learning data creation device that creates learning data used for object recognition.

  For object recognition as a technology for recognizing objects such as objects contained in images, prepare feature quantities for each object (representation of image features such as color scheme, texture, shape, etc.) It is necessary to keep. The feature amount of the object is obtained by preparing a large amount of image data for learning and machine learning the image.

  Therefore, in order to perform highly accurate object recognition, it is necessary to prepare a large amount of learning data that correctly represents the object. Since learning data is generally generated by a person visually confirming the contents of an image and labeling an image including an object, human labor is required.

  Further, in recent years, web search has become widespread and it has become possible to collect a large amount of data from the web, so that related images can be collected by performing a web search using a keyword representing an object. became.

  However, since a search index in web search is generated using a keyword included in a web page, an object representing the keyword is not always included in the searched image.

  Even if an object is included, there may be cases where the shooting state is not suitable for the learning data, such as when the object is small due to shooting from a distance, or when the lighting is insufficient. . For this reason, it took a great deal of time and effort to manually sort out images from images collected by web search or to cut out images from images.

  As a technique to reduce the time and effort required to judge whether or not the learning data is correct by such a person, the user can select a correct case from the clusters obtained by dividing the image into multiple regions and clustering the region images. A technique for creating learning data (image dictionary) based on this selection is known (see Patent Document 1).

JP 2009-282660 A

  However, even in the technique of Patent Document 1, it is necessary to make a judgment by a person to select a positive case from clusters, and the judgment becomes complicated and complicated as the number of clusters increases.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to automatically create learning data from image data collected from the web without human intervention.

  In order to achieve the above object, the first invention provides a learning data creating apparatus for creating learning data for object recognition from a plurality of image data collected by a web search based on a keyword, the plurality of collected image data The region images are classified into predetermined clusters based on the feature amount of the region image detected from the image, and a cluster to which the region image of each image data belongs and the number of region images belonging to the cluster are generated for each image data. And a non-common area specifying means for specifying a cluster with a small distribution of the area image from a classification state of the area image with respect to each cluster based on the number of area images belonging to each cluster; Learning data is created by removing the region image belonging to the cluster from the image data in which the region image is detected. Is characterized by comprising a learning data creation means for, the.

  According to the first invention, a cluster having a small distribution of region images is identified from the classification state of the region image with respect to the cluster based on the number of region images belonging to each cluster, and the region image belonging to the cluster is removed from the image data. To create learning data. For this reason, since image areas having no commonality between images are removed, image areas having commonality between images remain in the learning data. Among the image data collected by the keyword search, it is presumed that an image area having more common includes an object representing the keyword. Therefore, learning data can be automatically created from image data collected from the web without human intervention.

  Further, the non-common area specifying means in the second aspect of the invention is characterized in that, based on the number of image data from which the area image classified into the cluster is detected, out of the clusters into which the area image is classified. It is characterized by identifying clusters with low distribution.

  According to the second aspect of the invention, since a cluster having a small distribution of region images is specified based on the number of image data from which the region images classified into the clusters are detected, it is common depending on which cluster the image is classified into. Sex image areas can be identified.

  Further, the third invention is based on a ratio of the number of area images of the image data belonging to the cluster specified by the non-common area specifying means to the number of area images detected from the image data. Is further provided with quality determining means for determining whether or not the image data is suitable for the learning data, and the learning data creating means is determined to be suitable for the learning data by the quality determining means. The learning data is created by removing the region image from the image data.

  According to the third invention, since the learning data is created depending on whether or not the image data when the region image is removed is suitable for the learning data, it is possible to create high-quality learning data suitable for object recognition. it can.

  Further, the learning data creating means in the fourth invention removes the region image from the outside with respect to the image data based on the detection position of the region image classified into the specified cluster with respect to the image data. It is characterized by that.

  According to the fourth invention, since the region image is removed from the outside of the image based on the detection position of the region image classified into the identified cluster, the object is lacking in the image data created as the learning data. It can be prevented from occurring.

  According to the present invention, learning data can be automatically created from image data collected from the web without human intervention.

The block diagram which shows the function structure of the learning data creation apparatus which concerns on this invention. The flowchart of a feature vector generation process. The figure which shows the mode of the extraction of the area image from image data, and the mapping to a visual keyword. The flowchart of learning data creation processing. The conceptual diagram for demonstrating specification etc. of a non-common area | region. The figure which shows the example of creation of learning data. The figure for demonstrating the other Example of the removal of a non-common area | region.

[Configuration of image search device]
Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a functional block diagram of a learning data creation apparatus 1 to which the present invention is applied. The learning data creation device 1 is connected to the Internet connected via a communication network, and can collect image data from the web via the Internet. Learning data used for object recognition is created by cutting out and selecting an image area including the object from the collected data.

  The learning data creation device 1 in the present embodiment identifies non-common image areas across image data, determines that no common object is included in the identified image areas, and creates learning data. . A visual keyword technique is used to identify the non-common image area.

  A visual keyword refers to an image based on a feature amount obtained from image regions (hereinafter referred to as “region images” and “partial images” as appropriate) that consider each image as a set of a plurality of fine image regions. This is a technique for generating an index (feature vector), and can be said to be an application of a text technique for obtaining a feature amount of a sentence from a keyword in the text.

  For this reason, in the visual keyword, a region vector (visual segment) in the image is treated as a keyword, so that a fine partial region of the image can be analyzed and a feature vector representing one image can be generated. . In addition, text technology that performs document analysis as a set of words (keywords) (transposition index, vector space model, word appearance frequency, etc.) can be applied to the feature vector of the image. Can be realized.

As reference technical literature on image search using visual keywords,
Sivic and Zisserman: “Efficient visual search for objects in videos”, Proceedings of the IEEE, Vol.96, No.4, pp.548-566, Apr 2008.
・ Yang and Hauptmann: “A text categorization approach to video scene classification using keypoint features”, Carnegie Mellon University Technical Report, pp. 25, Oct 2006.
・ Jiang and Ngo: “Bag-of-visual-words expansion using visual relatedness for video indexing”, Proc. 31 ^st ACM SIGIR Conf., Pp.769-770, Jul 2008.
・ Jiang, Ngo, and Yang: “Towards optimal bag-of-features for object categorization and semantic video retrieval”, Proc. 6th ACM CIVR Conf., Pp.494-501, Jul. 2007.
・ Yang, Jiang, Hauptmann, and Ngo: “Evaluating bag-of-visual-words representations in scene classification”, Proc. 15 ^th ACM MM Conf., Workshop on MMIR, pp.197-206, Sep. 2007.
Etc.

  As shown in FIG. 1, the learning data creation device 1 includes an image collection unit 10, an image DB (database) 15, a visual keyword generation unit 20, a visual keyword DB 25, a feature vector generation unit 30, an area management DB 35, a feature vector DB 40, The non-common area specifying unit 50, the quality determining unit 60, the learning data creating unit 70, and the learning data DB 75 are configured.

  These functional units are configured by so-called computers, and are associated with a CPU (Central Processing Unit) as an arithmetic / control device, a RAM (Random Access Memory) and a ROM (Read Only Memory) as a storage medium, a communication interface, and the like. It is realized with.

  The image collection unit 10 is a functional unit that collects image data from the web via the Internet. The image collection unit 10 acquires a web page associated with the keyword by, for example, transmitting a predetermined keyword to the search engine. Then, the image data included in the web page is extracted, and the keyword and the image data are associated with each other and stored in the image DB 15.

  Further, the search engine may be an image search engine that uses image data as a search target. In this case, image data of a search result returned in response to transmission of a keyword is received and stored in the image DB 15. To do.

  The image DB 15 is a database for accumulating and storing image data collected by the image collecting unit 10, and stores keywords, image IDs, and image data in association with each other as shown in FIG. The image ID is identification information for uniquely identifying each image data, and is assigned by the image collection unit 10 when storing a keyword and image data.

  When generating the feature vector of the image data, the visual keyword generation unit 20 generates a classification (cluster) for mapping the region image in the image. The visual keyword generation unit 20 extracts a plurality of region images from images used for image search and image data prepared in advance for learning, and clusters these images based on the feature values of the region images. As a standard method of clustering, k-means, Hierarchical Agglomerative Clustering (HAC) or the like is used.

  The feature vector generation unit 30 to be described later generates a feature vector by mapping (classifying) the region image detected from the image to a cluster formed by the clustering of the visual keyword generation unit 20. This cluster is referred to as a “visual keyword” as a feature amount space for expressing an image as a collection of visual keywords.

  The visual keyword DB 25 includes a visual keyword ID (VKID) for identifying a cluster formed by clustering of the visual keyword generation unit 20, and a center coordinate that is a coordinate of a center point in the feature amount space (multidimensional space) of the cluster. , A database that stores the radius indicating the cluster range in association with each other.

  The center coordinate is a value indicating an average value of feature amounts of images belonging to each cluster, and is represented by multidimensional coordinates on the feature amount space. The radius is obtained, for example, by the distance from the image farthest from the center coordinate among the images belonging to the cluster.

  The feature vector generation unit 30 extracts a region image from the image data, performs a feature vector generation process (see FIG. 2) that generates a feature vector based on the feature amount of the region image, and obtains a feature vector of each image data. Generate. The feature vector generation process will be described later.

  The area management DB 35 is a database that stores a correspondence relationship between the area image detected from each image data by the feature vector generation unit 30, the visual keyword of the mapping destination, and the image data of the area image detection source, As shown in FIG. 1, an image ID, a region ID, and a VKID are stored in association with each other.

  The feature vector DB 40 is a database that stores the feature vectors generated by the feature vector generation unit 30 in association with each image. As shown in FIG. 1, the feature vector DB 40 stores an image ID and a region image for each visual keyword that becomes a feature vector. The appearance frequency is stored in association with each other.

  Here, the feature vector generation processing will be described with reference to the flowchart of FIG. 2 and the conceptual diagram of FIG.

  First, the feature vector generation unit 30 reads the image data stored in the image DB 15 and detects a plurality of region images from the image data (step S11). As a method for detecting the region image, there are a method for detecting a characteristic region (characteristic region) in the image and a method for detecting the region image by dividing the image into predetermined regions.

As a technique for detecting feature regions,
・ Harris-affine
・ Hessian-affine
・ Maximally stable extremal regions (MSER)
・ Difference of Gaussians (DoG)
・ Laplacian of Gaussian (LoG)
・ Determinant of Hessian (DoH)
Etc.

  The feature region detection technology is published in “Local Invariant Feature Detectors: A Survey” (Foundations and Trends in Computer Graphics and Vision, Vol.3, No.3, pp.177-280, 2007.). Any known technique can be adopted as appropriate.

  In addition, as a method of detecting an image by dividing it into predetermined areas, for example, the image is divided into predetermined M × N blocks, or divided so that the size of the divided blocks becomes predetermined m × n pixels. There is a technique to do. For example, when an image is divided into 10 × 10 blocks, if the size of the image is 640 × 480 pixels, the size of one block is 64 × 48 pixels.

  3 shows an example in which an image is divided into predetermined areas. The 0001 image is divided into 7 × 6 blocks. No. For images of 0002, 5 × 7 blocks, No. The 0003 image is divided into 6 × 6 blocks. In the illustrated example, it is divided into several blocks for simplification of explanation, but is divided into tens of thousands of blocks.

  Next, the feature vector generation unit 30 calculates the feature amount of the detected region image (step S12). When the feature region is extracted, a local feature amount resistant to affine transformation such as scale change, rotation, and angle change is extracted. Examples of the local feature amount include the following.

・ SIFT
・ Gradient location and orientation histogram
・ Shape context
・ PCA-SIFT
・ Spin images
・ Steerable filters
・ Differential inverters
・ Complex filters
・ Moment inviteants

  The extraction of local features is published in “A performance evaluation of local descriptors” (IEEE Trans. Pattern Analysis and Machine Intelligence, Vol.27, No.10, pp.1615-1630, 2005.) A known technique can be adopted as appropriate.

  Since the feature vector generated based on the feature amount extracted from the feature region is generated from the feature region where the object (object) is highly likely to exist, it is effective as an index indicating the feature of the object in the image.

  In addition, when a region image is extracted by region division, an image feature amount expressed by quantifying features of each image such as an image color scheme, texture, and shape is used. Since the feature vector generated from the feature amount of the region image detected by the region division is generated from each part constituting the image, it is effective as an index indicating the overall configuration of the image.

  Then, the feature vector generation unit 30 maps (classifies) the plurality of region images detected from the image data to visual keywords based on the feature amounts of the region images (step S13). The mapping to the visual keyword is performed by selecting the visual keyword having the closest distance based on the distance in the feature amount space between the center point of each visual keyword (cluster) and the feature amount of the region image.

  In the example of FIG. 3, the area images T1, T3 to T6 detected from the image with the image ID “0001” are mapped to the visual keyword VK1, and the area image T2 is mapped to the visual keyword VK2. Further, the region images T12 to T14 detected from the image with the image ID “0002” are mapped to the visual keyword VK3. Further, the area image T11 of the image with the image ID “0002” and the area image T21 of the image with the image ID “0003” are mapped to the visual keyword VK4.

  When each area image is mapped to a visual keyword (cluster), the feature vector generation unit 30 counts the appearance frequency of the area image for each visual keyword, and displays it in a multidimensional manner according to the appearance frequency of the area image for each visual keyword. The generated feature vector is generated and stored in the feature vector DB 40 (step S14).

  For example, in the case of the image “0001” in FIG. 3, the appearance frequency of the region image detected from the image is “5” for the visual keyword VK1, “1” for the visual keyword VK2, and “0” for the visual keyword VK3. Become. A feature vector having the appearance frequency for the plurality of visual keywords as a vector element is generated.

  The feature vector generation unit 30 assigns a region ID to the region image detected from the image data, and stores the VKID of the visual keyword mapping the region image in the region management DB 35 in association with the image ID and the region ID. The area ID may be an XY coordinate in the image, or may be a row number / column number when the area is divided.

  The non-common area specifying unit 50 uses the feature vector generated by the feature vector generating unit 30 to specify an image area that is not common across the image data. Although details will be described later, in brief, based on a feature vector generated for each of a plurality of image data (hereinafter also referred to as “original images”) from which the region image is detected, the original image is A mapped visual keyword, that is, a distribution state of the original image in the same feature amount space is calculated, and an image region that is not common between the images is determined based on the distribution state.

  The quality determination unit 60 determines the quality of whether or not the removed image data is suitable as learning data when the non-common image region determined by the non-common region specifying unit 50 is removed from the original image. To do. This is because it is preferable that the learning data used for object recognition sufficiently express the characteristics of the object with a certain level of quality, and the quality determination unit 60 is an area when the non-common area is removed from the original image. The quality is determined based on the mapping state of the image to the visual keyword.

  The learning data creation unit 70 removes the region image specified as the non-common region by the non-common region specifying unit 50 from the image data determined to be suitable for the learning data by the quality determination unit 60, and the learning data Is stored in the learning data DB 75.

  The learning data DB 75 is a database that stores keywords and learning data in association with each other. The learning data DB 75 associates the learning data created by the learning data creation unit 70 with the image data from which the learning data is created. Are stored in association with each other.

  The learning data stored in the learning data DB 75 is used by the object recognition device to learn the characteristics of the object represented by each keyword. The learning method, the feature amount extraction method, and the like performed by the object recognition device do not depend on the algorithm of the learning data creation device 1, but are arbitrarily set by the object recognition device.

  In the present embodiment, the image data to be stored as learning data is selected by the quality determination of the quality determination unit 60, but it is specified as non-common by the non-common area specifying unit 50 without performing this quality determination. Learning data may be created by removing the image area and registered in the learning data DB 75.

[Learning data creation process]
Next, learning data creation processing executed by the non-common area specifying unit 50, the quality determining unit 60, and the learning data creating unit 70 will be described using the flowchart of FIG. 4 and the conceptual diagram of FIG.

  First, the non-common area specifying unit 50 includes the number of visual keywords (number of VK assignments) for each image obtained by mapping the area images of each image, and the number of original images distributed to each visual keyword (number of original image distributions). And calculate.

  Specifically, for a plurality of image data associated with the same keyword, the number of visual keywords mapped to the region image detected from each image is calculated for each image (step S21). For example, as shown in FIG. 5, there are three mapping destinations of the area image detected from the image IMG1, VK1, VK2, and VK4, and the VK allocation number is calculated as ‘3’.

  Further, the number of original images in which area images are mapped to each visual keyword is calculated for each visual keyword as the original image distribution number (step S22). For example, in FIG. 5, since the region image is mapped from the three original images IMG1, IMG2, and IMG3 to the visual keyword VK1, the number of original image distributions is calculated as “3”.

  Next, the non-common area specifying unit 50 selects one visual keyword (step S23), and determines whether or not the number of original image distributions of the visual keyword is less than a predetermined threshold (step S24). If it is determined that the number of original image distributions is less than the predetermined threshold (step S24; Yes), the region image mapped to the visual keyword is specified as a non-common region (step S25).

  The non-common area specifying unit 50 determines whether or not the processing of steps S23 to S25 has been performed for all visual keywords (step S26), and if there is an unprocessed visual keyword, the processing proceeds to step S23.

  For example, if the threshold is set to '3' and the selected visual keyword is VK1, the feature amount space represented by this visual keyword VK1 is determined to be a common area. Further, as shown in FIG. 5, since the visual keyword VK2 is less than the threshold value, it is determined as a non-common area.

  In this way, by using the feature vector generated using the visual keyword, it is possible to specify a cluster on the feature amount space, which is a visual keyword that is not common between the images.

  The learning data creation apparatus 1 of the present embodiment removes an image area belonging to this non-common visual keyword from the original image, thereby leaving a common image area between the images. Since the image area common to the images is an image area having a common feature amount in the image set obtained by the keyword search, it can be said that the object represented by the keyword is included. Therefore, an image including an object can be generated by removing the non-common area from the original image.

  Note that the threshold value used for specifying whether or not it is a non-common area in step S24 may be a constant as described above, or the number of images P% (for example, 10%) of the total number of images collected based on the keyword. It is good also as setting dynamically. It is also possible to select and set the largest one of a constant or P% of the total number of images.

  If the non-common area specifying unit 50 determines that all visual keywords have been processed in step S26 (step S26; Yes), the quality determination unit 60 selects the original images one by one (step S27), Steps S28 to S31 are performed.

  First, it is determined whether or not the image when the image area belonging to the visual keyword identified as the non-common area from the selected original image is removed from the original image satisfies the quality condition (step S28).

  Specifically, with respect to the total number of visual keywords (total number of remaining VKs) when visual keywords that have become non-common regions are removed from all images, the distribution of images to visual keywords when non-common regions are removed ( If the ratio of (number of remaining VK) is equal to or greater than a predetermined value, it is determined that the quality condition is satisfied. The ratio of the distribution due to the removal of the non-common area is obtained by the following equation.

  Ratio of distribution of each image = number of remaining VKs / total number of remaining VKs

  The number of remaining VKs is the number of visual keywords remaining after mapping the image area even when the image area mapped to the visual keyword of the non-common area is removed from the original image, and is calculated for each image. In FIG. 5, it is the number of visual keywords excluding the visual keywords surrounded by a broken line. In the image IMG1, since the visual keywords VK1 and VK4 remain, the remaining number of VKs = 2. In the image IMG4, only the visual keyword VK4 is mapped and remains, so the number of remaining VKs = 1.

  The total number of remaining VKs is the total number of remaining visual keywords even when all the images of the visual keywords in the non-common area are mapped and removed from the remaining visual keywords, and is obtained for the entire collected image. In FIG. 5, there are six visual keywords to which all images are mapped, and among them, there are four visual keywords specified as non-common areas, so the total number of remaining VKs = 2.

  When the ratio of the remaining number of VKs of each image to the total number of remaining VKs is equal to or less than a predetermined threshold (for example, 0.5), it is determined that the image does not satisfy the quality as learning data. This means that the number of visual keywords expressing objects by removing non-common areas is relatively small with respect to the entire set (total number of remaining VKs).

  In other words, if there are fewer visual keywords (clusters in the feature space) remaining after removing the non-common area, it is assumed that the features sufficient to represent the object have been deleted by the removal, and the learning data Is determined to be inappropriate.

  Further, for example, if an image area that is not common to an image in which an object has been photographed or a part of the object has been photographed is removed, the remaining image becomes small and cannot be used as learning data. Even in such a case, learning data suitable for learning can be selected and registered by the above-described quality determination.

  Note that the threshold value in this quality determination may be a constant as described above, or may be an average value of the distribution ratio of each image. Further, it is possible to select and set the maximum of either the constant or the average value of the ratio of the image distribution.

  When it is determined that the image does not satisfy the quality condition (Step S28; No), the quality determination unit 60 deletes the image from the image DB 15 (Step S30), and selects the next image (Step S31 → S27).

  When it is determined that the image satisfies the quality condition (step S28; Yes), learning data is created by removing the image area mapped to the visual keyword of the non-common area from the image, and learning data The keyword is stored in the DB 75 together with the keyword (step S29).

  As described above, for example, “land” is removed as shown in FIG. 6 by removing the image area that is identified as non-common between the images based on the feature vector generated from each image using the visual keyword. Among the images collected with the keyword “mark”, an image region (region surrounded by a broken line) representing the object “landmark” can be cut out to create learning data suitable for object recognition.

  In addition, when an image area specified as non-common is removed as in the image IMG4 of FIG. 6, if the quality as learning data is not satisfied, the image data is not registered as learning data. Highly accurate learning data can be created.

  The embodiment described above is an example to which the present invention is applied, and the design may be changed as appropriate without departing from the object of the present invention. Hereinafter, modifications of the present invention will be described.

[Change of distribution index to visual keywords]
First, as the degree of distribution of the image of the visual keyword, in the above example, the number of original image distributions is calculated, but the number of image areas mapped to each visual keyword is calculated to specify the non-common area. Also good.

  Specifically, in step S21 of FIG. 4, instead of calculating the VK allocation number, the detection number of area images detected from each image is calculated as the detection number for each image. In step S22, instead of calculating the number of original image distributions, the total number of area images mapped to each visual keyword is calculated as the number of VK mappings. In step S25, a visual keyword having a VK mapping number less than a predetermined threshold is specified as a non-common area.

  For example, in FIG. 5, 15 area images from image IMG1, 11 images from IMG2, and 7 area images from image IMG3 are mapped to visual keyword VK1, and therefore, 33 obtained by adding them is calculated as the number of VK mappings. . When the threshold is set to 20, the visual keywords VK2, VK5, and VK6 are specified as non-common areas in FIG.

  Further, in step S28 for determining quality as learning data, when non-common areas are removed with respect to the number of remaining image areas (total number of remaining areas) when visual keywords that have become non-common areas are removed from all images. If the ratio of the distribution (number of remaining detections) of the image area to the visual keyword of each image is equal to or greater than a predetermined value, it is determined that the quality condition is satisfied. The ratio of the distribution due to the removal of the non-common area is obtained by the following equation.

  Distribution ratio of each image = number of remaining detections / total number of remaining areas

  The total number of remaining areas is the number of image areas remaining mapped to each visual keyword even when the image area mapped to the visual keyword of the non-common area is removed from the original image, and is calculated for each image. In FIG. 5, the total number “88” of image areas mapped to the visual keywords VK1, VK3, and VK4 is calculated as the total number of remaining areas.

  The number of remaining detections is the number of area images remaining in each image even when the area image of the visual keyword in the non-common area is removed, and is calculated for each image. In FIG. 5, for the image IMG1, the area images mapped to the visual keywords VK1, VK3, and VK4 other than the non-common areas are calculated as 24 images.

  If the ratio of the number of remaining detections of each image to the total number of remaining areas is a predetermined threshold (for example, 0.5) or less, it is determined that the image does not satisfy the quality as learning data. In this way, it is possible to create higher quality learning data by using the number of area images mapped to each visual keyword to specify the visual keywords that are non-specific areas and to determine the quality as learning data. Can do.

[Removal of image area mapped to non-common area]
In the above example, the learning data is created by removing the region image belonging to the visual keyword specified as the non-common region from the original image. However, the visual keyword specified as the non-common region is described. An area to be actually removed may be obtained based on the position of the area image belonging to the original image in the original image.

  Specifically, in the image IMG5 shown in FIG. 7, it is assumed that the image area is divided into areas indicated by broken lines and the shaded area is an image area mapped to a visual keyword specified as a non-common area.

  As shown in FIG. 7, the image area (white area) that becomes the common area can be recognized from the position of the image area mapped to the non-common area. An image area is extracted, and learning data is cut out so as to include the image area.

  That is, in the image IMG5, the image areas P1 to P4 are extracted as the outer edges of the image areas that are common areas, and the frame F including the image areas P1 to P4 is extracted. By removing the image area outside the frame F from the image IMG, an image inside the frame F is created as learning data.

  The feature amount used for object recognition using the learning data created by the learning data creation device 1 of the present embodiment differs depending on the object recognition engine. Therefore, by creating the learning data with a sufficient size including the object, it is possible to accurately extract the characteristics of the object in the learning process.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

1 learning data creation device 10 image collection unit 15 image DB
20 Visual keyword generator 25 Visual keyword DB
30 feature vector generation unit 35 area management DB
40 Feature vector DB
50 Non-common area specifying unit 60 Quality determining unit 70 Learning data creating unit 75 Learning data DB

Claims (6)

In a learning data creation device for creating learning data for object recognition from a plurality of image data collected by web search based on keywords,
The area image is classified into a predetermined cluster based on the feature amount of the area image detected from the collected plurality of image data, the cluster to which the area image of each image data belongs, and the number of area images belonging to the cluster And classifying means for generating the image data for each image data;
Non-common area specifying means for specifying a cluster with a small distribution of the area image from the classification state of the area image for each cluster based on the number of area images belonging to each cluster;
Learning data creating means for creating learning data by removing the region image belonging to the identified cluster from the image data in which the region image is detected;
A learning data creation device comprising: The non-common area specifying means includes
2. The cluster in which the distribution of the region image is small is identified based on the number of image data from which the region image classified into the cluster is detected, among the clusters into which the region image is classified. The learning data creation device described in 1. Based on the ratio of the number of area images of the image data belonging to the cluster specified by the non-common area specifying means to the number of area images detected from each image data, the image data when the area image is removed is Further comprising quality judging means for judging whether or not the learning data is suitable;
The learning data creation means includes
3. The learning data generating apparatus according to claim 1, wherein the learning data is generated by removing the region image from image data determined to be suitable for the learning data by the quality determination unit. . The learning data creation means includes
The area image is removed from the outside of the image data based on a detection position of the area image classified into the identified cluster with respect to the image data. The learning data creation device described. In a learning data creation method in which a computer creates learning data for object recognition from a plurality of image data collected by web search based on keywords,
The area image is classified into a predetermined cluster based on the feature amount of the area image detected from the collected plurality of image data, the cluster to which the area image of each image data belongs, and the number of area images belonging to the cluster A cluster process for generating the image data for each image data;
A non-common area specifying step for specifying a cluster with a small distribution of the area image from the classification state of the area image for each cluster based on the number of area images belonging to each cluster;
A learning data creating step of creating learning data by removing the region image belonging to the identified cluster from the image data in which the region image is detected;
The learning data creation method, wherein the computer performs the above.   A program for causing a computer to execute the learning data creation method according to claim 5.

Images