JP2008022441A - Image processing apparatus and method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a video analysis technology for more efficiently managing and handling video contents. <P>SOLUTION: A HSV histogram model storage section 51 stores a HSV histogram model for denoting image characteristics of a region of interest obtained as a result of analysis applied to a plurality of images each including the region of interest according to a prescribed analysis method. A HSV component peak detection section 53 applies analysis according to the prescribed analysis method to a processing object image that is an image newly received via a down-sampling section 52. A region of interest detection section 55 detects the region of interest included in the processing object image on the basis of an analysis result by the HSV component peak detection section 53 and a range of the region of interest determined by using the HSV histogram model. The technology above can be applied to fields handling video contents. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an image processing apparatus, method, and program, and in particular, it has become possible to provide a video analysis technique for managing and handling video content more efficiently so that the video content can be handled with a human intuitive feeling. The present invention relates to an image processing apparatus and method, and a program.

  Due to the rapid development of digitization technology in the fields of Internet technology, broadcasting, communication, entertainment, medical care, education, etc., there is an increasing need for easy and high-speed access to only necessary information from a huge amount of multimedia information. . Among such needs, research and development of image recognition technology for semantic extraction and analysis of video content, especially from the need to efficiently manage and handle large amounts of video content stored on networks and home servers Has been actively performed.

  As a concrete example, for example, while hard disk recording and digital home appliances are widespread, to make it possible to enjoy according to a new lifestyle such as watching only what you want to watch, enjoying playback while recording, etc. "Similar image search technology", digital cameras and digital video cameras, and the spread of the Internet, easily edit various images and video, "digest video creation to create and save your original digest video" Research and development such as “technology” has been performed (see, for example, Patent Documents 1 and 2).

  In other words, it is possible to enjoy according to the new lifestyle described above, or to search for a desired image from video content as a need for creating a digest video, for example, to detect a highlight scene from video content, There is a growing need for image search.

In order to respond to the needs for such image retrieval, image retrieval techniques using keywords have been put into practical use. In addition, research was also conducted on Browsing technology that allows viewers to freely view video content and technology for creating meaningful “video minutes” in visual communication.
JP 2002-344872A JP 2006-54622 A

  However, the conventional image search technology using keywords is time-consuming to assign the necessary keywords, and there is a tendency that a large number of images of the keywords exist, so that actually satisfactory search results cannot be obtained. Is the current situation. In addition, search technology that automatically extracts meta information in video and finds desired image content based on the meta information, technology that automatically creates “video minutes”, browsing technology, etc. Although much research has been done, it is not easy to extract the necessary meta information with high accuracy, so that actually available video search technology, digest creation technology, and browsing technology have not been established.

  In other words, there is a demand for the realization of video analysis technology for more efficiently managing and handling video content so that the video content can be handled with a human intuitive feeling, rather than image search using keywords as described above. However, it is a situation where the request is not fully met.

  The present invention has been made in view of such a situation, and can provide a video analysis technique for managing and handling video content more efficiently so that the video content can be handled with a human intuitive feeling. To do.

  An image processing apparatus according to one aspect of the present invention is an image processing apparatus that performs image processing on an image obtained as a result of photographing a real-world subject, and has one or more real-world backgrounds for the real-world subject. An analysis according to a predetermined analysis technique is performed in advance on a plurality of images including a target area corresponding to the target background as a target background and the target area corresponding to the target background, and obtained as a result of the analysis. A holding unit that holds image characteristic information indicating an image characteristic of a region of interest, and an analysis in which the newly input image is processed as an image to be processed and analysis according to the predetermined analysis method is performed on the processing target image Means, and detection means for detecting the region of interest included in the processing target image based on the analysis result of the analysis means and the image characteristic information held in the holding means.

  The predetermined analysis method is a method of analyzing a 3D HSV (3-Dimensions Hue Saturation Value) histogram of an image, and the image characteristic information is information indicating characteristics of the 3D HSV histogram of an image including the region of interest. is there.

  The image characteristic information is information indicating each dispersion range of H, S, and V that can be taken by an image including the region of interest, and the analysis unit calculates the 3D HSV histogram for the processing target image, Peaks of each component of H, S, V are detected from the 3D HSV histogram, and each peak detected by the analysis unit is held in the holding unit as each central value of H, S, V. The image processing apparatus further includes a determination unit that determines each range of H, S, and V having the respective dispersion widths of H, S, and V as a region of interest range, and the detection unit includes the process Of the target image, an area composed of pixels having H, S, and V component values within the area of interest determined by the determining means is detected as the area of interest.

  The image processing apparatus further includes an object feature extracting unit that extracts an object feature in the region of interest detected by the detecting unit in the processing target image or a region in contact with the target region.

  The object feature extraction unit detects a boundary of the attention area from the processing target image, and extracts information indicating the detection result as one of the object features.

  The processing target image is an image obtained as a result of shooting a soccer game, the attention area is an area corresponding to a real-world soccer play field, and the object feature extraction means further includes The area corresponding to the goal in the real world of soccer from the processing target image is an area surrounded by two line segments in contact with the boundary of the attention area and one line segment connecting the two line segments. Then, an area that satisfies the stipulation of the size of the goal is detected, and information indicating the detection result is extracted as one of the object features.

  The object feature extraction means further detects a region corresponding to a corner position in the real world of soccer based on the boundary of the region of interest, and extracts information indicating the detection result as one of the object features.

  An image processing method according to an aspect of the present invention is an image processing method of an image processing apparatus that performs image processing on an image obtained as a result of photographing a real-world subject, and includes one or more real-world subjects. An analysis according to a predetermined analysis method is performed in advance on a plurality of images including an attention background corresponding to the attention background as a predetermined one of the world backgrounds, Image characteristic information indicating the image characteristics of the attention area obtained as a result is held in the image processing apparatus, and a newly input image is used as a processing target image, and analysis according to the predetermined analysis method is performed as the processing target. Applying to the image and detecting the region of interest included in the processing target image based on the analysis result and the stored image characteristic information.

  A program according to one aspect of the present invention is a program corresponding to the above-described image processing method according to one aspect of the present invention.

  In the image processing apparatus, method, and program according to one aspect of the present invention, the following image processing is performed on an image obtained as a result of photographing a subject in the real world. That is, according to a predetermined analysis method, a predetermined one of one or more real world backgrounds with respect to a real-world subject is set as a target background, and a plurality of images including a target region corresponding to the target background are targeted. When an image is input in a state where image characteristics information indicating the image characteristics of the region of interest obtained as a result of the analysis is stored in the image processing apparatus or the like, a new input is performed. Using the image as a processing target image, analysis according to the predetermined analysis method is performed on the processing target image, and based on the analysis result and the stored image characteristic information, the processing target image The region of interest included in is detected.

  As described above, according to one aspect of the present invention, a video analysis technique can be provided. In particular, it is possible to provide video analysis technology for managing and handling video content more efficiently so that the video content can be handled with a human intuitive feeling.

  Embodiments of the present invention will be described below. Correspondences between the configuration requirements of the present invention and the embodiments described in the detailed description of the present invention are exemplified as follows. This description is to confirm that the embodiments supporting the present invention are described in the detailed description of the invention. Accordingly, although there are embodiments that are described in the detailed description of the invention but are not described here as embodiments corresponding to the constituent elements of the present invention, It does not mean that the embodiment does not correspond to the configuration requirements. Conversely, even if an embodiment is described here as corresponding to a configuration requirement, that means that the embodiment does not correspond to a configuration requirement other than the configuration requirement. It's not something to do.

  Further, this description does not mean that all the inventions corresponding to the specific examples described in the embodiments of the invention are described in the claims. In other words, this description is an invention corresponding to the specific example described in the embodiment of the invention, and the existence of an invention not described in the claims of this application, that is, in the future, a divisional application will be made. Nor does it deny the existence of an invention added by amendment.

An image processing apparatus according to one aspect of the present invention (for example, the image processing apparatus in FIG. 1)
An image processing apparatus that performs image processing on an image obtained as a result of photographing a real-world subject,
A plurality of images (for example, images 71 to 74 in FIG. 14) including a target region corresponding to the target background, with a predetermined one of one or more real-world backgrounds for a real-world subject as a target background. As a holding means (of the attention area extraction unit 11 in FIG. 1) in which analysis according to a predetermined analysis method is performed in advance and image characteristic information indicating the image characteristics of the attention area obtained as a result of the analysis is held. HSV histogram model holding unit 51) of FIG.
Analyzing means (for example, the HSV component of FIG. 12 in the attention area extracting unit 11 of FIG. 1) that performs the analysis according to the predetermined analysis method on the processing target image using the newly input image as the processing target image. Peak detector 53),
The region of interest included in the processing target image is detected based on the analysis result by the analyzing unit and the image characteristic information held in the holding unit (for example, in the example of FIG. Detection means (for example, the attention area detection section 55 in FIG. 12 of the attention area extraction section 11 in FIG. 1).

The predetermined analysis method is a method of analyzing a 3D HSV (3-Dimensions Hue Saturation Value) histogram (for example, see graphs 71 to 73 in FIG. 13) of an image,
The image characteristic information is information indicating characteristics of the 3D HSV histogram of an image including the region of interest.

The image characteristic information is information indicating each dispersion range of H, S, V that can be taken by the image including the region of interest,
The analysis means calculates the 3D HSV histogram for the processing target image, and detects the peak of each component of H, S, V from the 3D HSV histogram,
The respective peaks detected by the analyzing means are set as the respective center values of H, S, and V, and the respective dispersion widths of the H, S, and V held in the holding means are provided. , V as the attention area range, for example, the attention area determination section 54 in FIG. 12 of the attention area extraction section 11 in FIG.
The detection means detects, as the attention area, an area composed of pixels having H, S, and V component values within the attention area range determined by the determination means in the processing target image.

  The image processing apparatus further includes an object feature extraction unit (for example, the object feature extraction unit 12 in FIG. 1) that extracts an object feature in the region of interest detected by the detection unit in the processing target image or a region in contact therewith.

  The object feature extraction unit detects a boundary of the attention area (for example, boundary lines 123 and 124 in FIG. 18) from the processing target image, and extracts information indicating the detection result as one of the object features.

The processing target image is an image obtained as a result of shooting a soccer game,
The region of interest is a region corresponding to a real-world soccer playfield,
The object feature extracting means further includes two line segments in contact with the boundary of the region of interest and one line connecting the two line segments as an area corresponding to a goal in the real world of soccer from the processing target image. , A region that satisfies the stipulation of the size of the goal is detected, and information indicating the detection result (for example, goal gate 144 in FIG. 20) is extracted as one of the object features. .

  The object feature extraction means further detects an area corresponding to a corner position in the real world of soccer based on the boundary of the attention area (for example, boundary lines 162 and 163 of the play field area 161 in FIG. 21), Information indicating the detection result (for example, the corner position 164 in FIG. 21) is extracted as one of the object features.

An image processing method according to one aspect of the present invention (for example, a method corresponding to the image processing in FIG. 2)
An image processing method of an image processing apparatus (for example, the image processing apparatus in FIG. 1) that performs image processing on an image obtained as a result of photographing a real-world subject,
Analysis according to a predetermined analysis method with a predetermined one of one or more real world backgrounds for a real-world subject as a target background and a plurality of images including a target region corresponding to the target background. Is performed in advance, image characteristic information indicating the image characteristic of the region of interest obtained as a result of the analysis is held in the image processing device,
Using the newly input image as the processing target image, performing analysis according to the predetermined analysis method on the processing target image,
Based on the analysis result and the held image characteristic information, the step of detecting the attention area included in the processing target image (executed by the attention area extraction section 11 having the functional configuration of FIG. 12, for example) Step S1) of FIG. 2 is included.

  The program according to one aspect of the present invention is a program corresponding to the above-described image processing method according to one aspect of the present invention, and is executed by, for example, the computer of FIG.

  Thus, in the present invention having various aspects, video content is targeted. Here, content is widely generated by human creative activities. For example, movies, music, theatre, literary arts, photographs, comics, animations, computer games and other characters, figures, colors, sounds, actions or videos, or combinations thereof, or information related thereto are provided via an electronic computer. Is a content example. Of these, content mainly including video is referred to herein as video content. Video content can be broadly classified into still image content mainly including still images and moving image content mainly including moving images. In the following, the explanation will be made focusing on the moving image content.

  However, in this specification, so-called content data, that is, data generated by human creative activities, which can be processed by the device, for example, an electric signal, or fixed in a memory Are collectively referred to as contents without being particularly distinguished. That is, herein, an aggregate of data of each frame or field constituting a moving image is also referred to as moving image content.

  In the present invention, it is assumed that various types of image processing for moving image content are executed in units of access units. An access unit refers to a unit of a moving image such as a frame or a field, and specifically refers to, for example, the entire still image (frame or the like) or a part (field or the like) constituting the moving image. However, hereinafter, for the sake of simplicity, it is assumed that various image processing for moving image content is executed in units of frames.

  Some specific examples of images subjected to various image processing according to the present invention are shown in the drawings. However, from the viewpoint of copyright protection and the like, the actual image subject to image processing itself is not shown in the drawing, and an image with appropriate masking or a diagrammatic image Shown in the drawing. For example, the images 21 and 22 in FIG. 3 to be described later are not actual images subjected to image processing, but are displayed in the area corresponding to the fence of the spectator seat in the real world and the score display in the actual images. The corresponding area is an image after white masking. That is, FIG. 3 shows a specific example in which the attention area 24 described later is detected. However, the attention area 24 is not detected from the image 21 of the white mask shown in FIG. Note that the attention area 24 is detected from the actual image. Similarly, for example, the example of FIG. 4 to be described later is an example showing that the player's face 41 is detected from the frame 31, but image processing is applied to the frame 31 of FIG. The face 41 is not detected. Actually, the image processing is applied to the actual image data before the plotting of the frame 31 in FIG. 4, and the face as the actual image data corresponding to the face 41 is detected. ing.

  FIG. 1 shows a functional configuration example of an image processing apparatus to which the present invention is applied.

  The image processing apparatus in the example of FIG. 1 is configured to include an attention area extracting unit 11 to a highlight detecting unit 15.

  FIG. 2 is a flowchart for explaining an example of image processing executed by the image processing apparatus of the example of FIG.

  In step S1, the attention area extraction unit 11 extracts the attention area for each frame.

  Here, the attention area refers to a background to be noticed among one or more backgrounds in the real world among areas included in a still image (here, one frame) obtained as a result of photographing a subject in the real world. The area corresponding to.

  For example, consider a case where a video of a soccer game is used as video content. In this case, the real-world subject of the video content is a person related to a soccer game such as a player or an object related to a soccer game such as a ball, and the background of the real world is a play field where a soccer game is played. It can be broadly divided into (mainly grassy places) and other audience seats. In this case, considering the fact that it is a soccer game, the play field is a notable background. Therefore, the attention area extraction unit 11 can extract an area corresponding to the real-world play field (hereinafter referred to as a play field area) from among the areas included in the predetermined frame as the attention area.

  For example, the frame 21 in FIG. 3 is an image of a scene of a soccer game, and the area 23 is a play field area. The attention area extraction unit 11 extracts the play field area 23 as the attention area 24. That is, the image 22 is obtained as a result of the play field area being extracted as the attention area 24.

  Details will be described later with reference to FIGS. 12 to 16, but the attention area extraction unit 11 performs processing using an image feature model (for example, an HSV histogram model described later) of the play field area obtained in advance by learning. It is determined whether or not a play field area is included in a frame to be noted as a target (hereinafter referred to as a target frame). If it is determined that the play field area is included, the play field area is extracted as the attention area.

  The play field area is an example of an attention area when a video of sports such as soccer is adopted as moving image content. In addition, for example, a play field area of a broadcast program shot using a predetermined set in a studio in a broadcast station, for example. When the video is adopted as the moving image content, the set becomes a background of attention, and an area corresponding to the set can be detected as the attention area. Further, for example, when a video of a music concert is adopted as the moving image content, the stage becomes the attention background, and the area corresponding to the stage can be detected as the attention area. That is, any image region can be detected as a region of interest as long as the HSV histogram described later is characteristic as a result of capturing a real-world background that is characteristic of color.

  In step S2, the object feature extraction unit 12 extracts object features for each frame.

  Here, the object feature refers to image information indicating the feature of an object existing inside or adjacent to the region of interest. Further, it is assumed that a boundary line between the attention area and another area or an intersection of two or more boundary lines is also regarded as an object, and image information indicating the characteristics of the object is also included in the object characteristics. The image information is a broad concept including object position information in addition to object color information.

  For example, consider a case where a video of a soccer game is used as video content. In this case, since the play field area becomes the attention area, image information indicating the characteristics of a person, for example, a player or the like existing in the play field area can be detected. Specifically, in the example of FIG. 4, for example, the player's face 41 is detected from the frame 31 as one of the object features. Further, for example, the goal gate 42 is detected from the frame 32 in the example of FIG. 5 as the object feature information other than the person existing in the play field area or in the adjacent area. Further, for example, in the example of FIG. 6, the corner position 43 and the ball 44 are detected from the frame 33, and in the example of FIG. 7, the play field boundary 45 itself is also detected as one of the object features. In addition, although not shown, a white line or the like can also be detected as one of the object features.

  Details of the object feature detection will be described later with reference to FIGS.

  Returning to FIG. 2, in step S <b> 3, the shot cut detection unit 13 executes a process of detecting each shot cut in the moving image content (hereinafter referred to as a shot cut detection process). In other words, a shot between each shot cut is detected.

  For example, in the example of FIG. 8, between two frames described as “Frame25302” and “Frame25303” in the moving image content that is a video of a soccer game is detected as a shot cut. That is, in FIG. 8, each image described as “Frame25215” to “Frame25339” indicates each frame having the frame number. In this case, in the example of FIG. 8, frames numbered 25215 to 25302 are detected as the first shot, and frames numbered 25303 to 25339 are detected as the second shot different from the first shot. .

  Details of the shot cut detection process will be described later with reference to FIGS.

  Returning to FIG. 2, in step S4, the shot classification unit 14 classifies each shot in the moving image content into one of a plurality of predefined types (hereinafter referred to as a shot classification process). Execute.

  For example, in the example of FIG. 9, each image represents one shot, and each of eight consecutive shots of the video content that is a video of a soccer game is represented by an arrow indicating a time direction. , “Long”, “Close-up”, “Close-Up”, “Medium”, “Long”, “Medium”, “Long”, and “Medium”.

  The type of “Long” or the like will be described later with reference to FIG. Details of the shot classification process will be described later with reference to FIGS.

  Returning to FIG. 2, in step S5, the highlight detection unit 15 performs, for example, based on each object feature extracted in the process of step S2, the type of each shot classified in the process of step S4, or the like, for example, Executes processing to detect highlights from moving image content based on changes in feature values between frames belonging to each shot and time-series features between shots (hereinafter such processing is referred to as highlight detection processing). To do. Note that the highlight detection unit 15 can also create a digest based on the highlight, if necessary.

  For example, FIG. 10 and FIG. 11 show examples of information used when a series of scenes corresponding to corner kicks are highlighted and detected in video content that is a video of a soccer game. Yes.

  For example, FIG. 10 shows an example in which a highlight is detected by a change in feature amount between frames belonging to each shot. That is, in the example of FIG. 10, the feature amount (frame type, which will be described later) between frames belonging to a predetermined shot transitions as “Upper Corner” → “Unknown” → “Goal Box” → “Goal Box”. Judging from the results (statistical model for Layer 2 in FIG. 37 described later) of learning how various feature quantities have been learned so far from various videos of other soccer games, It means that it can be recognized as a method, and it has been detected as a highlight from the result of the authentication. The meaning of each feature quantity such as “Upper Corner” will be described later with reference to FIG.

  Further, for example, FIG. 11 shows an example in which highlights are detected as a time-series feature between shots depending on how shot types change. That is, in the example of FIG. 11, each shot type has transitioned from “Close-up” → “Medium” → “Long Shot” → “Close-up”. Judging from the results of learning with various videos of other soccer games (statistical model for Layer 1 in FIG. 37 described later), it can be recognized that the corner kick transition method is highlighted. Means that it was detected. Each type such as “Long” will be described later with reference to FIG.

  In this case, a corner kick highlight may be detected from the result of recognition of only one of the example of FIG. 10 or the example of FIG. 11, or both the examples of FIG. 10 and the example of FIG. In some cases, a corner kick highlight is detected by comprehensively judging the result, or both of the recognition results of the two and other information.

  Details of the highlight detection process will be described later with reference to FIGS.

  Hereinafter, the details of each of the attention area extracting unit 11 to the highlight detecting unit 15 constituting the image processing apparatus of FIG. 1 will be described individually in that order. At that time, in order to facilitate understanding of the present invention, when it is necessary to give specific examples of moving image content, soccer game images are employed in accordance with the examples of FIGS. And

  First, a detailed example of the attention area extraction unit 11 in the image processing apparatus of FIG. 1 will be described with reference to FIGS. 12 to 16.

  FIG. 12 shows a detailed example of the functional configuration of the attention area extraction unit 11.

  The attention area extraction unit 11 in the example of FIG. 12 is configured to include an HSV histogram model holding unit 51 to a post processing unit 56.

  The HSV histogram model holding unit 51 holds an HSV histogram model used for detecting a region of interest from the frame of interest, that is, a play field region in the present embodiment.

  Here, the HSV histogram model will be described.

  FIG. 13 is a graph obtained as a result of projecting a 3D HSV (3-Dimensions Hue Saturation Value) histogram to each 1D of representative sample images 61 to 63 corresponding to one scene of a soccer game video. Each of 71 to 73 is shown.

  As can be seen from the graphs 71 to 73, since the video of the soccer game includes a lot of play field regions mainly including the grass video, the HSV histogram characteristics (hereinafter referred to as 3D HSV histogram characteristics) are included. There is a certain relationship between frames.

  Therefore, the present inventor has made many images collected from various soccer game videos including the sample images 61 to 63 described above, that is, images in various time zones such as daytime, evening, and night, and sunny and cloudy. By analyzing each 3D HSV histogram feature such as images in various weather conditions such as rain, and applying statistical processing, the following statistical data results were obtained.

  That is, the variable range of HSV in the image including the noticeable play field area is H [22, 68], S [32, 255], V [64, 175], respectively, and the range width is the maximum. The results were H: 16, S: 128, V: 64, respectively. In other words, the average value of each component of HSV in the image including the play field region changes depending on the time zone of the soccer game, the climate, etc., but the variance width of each component is almost H: 16, S: 128 , V: turned out to be confined to 64. Therefore, it is possible to detect the play field region robustly by using such HSV features such as the dispersion widths of H, S, and V.

  Therefore, in the present embodiment, for example, such a dispersion range is held in the HSV histogram model holding unit 51 of FIG. 12 as an HSV histogram model.

  As training sample images for obtaining such HSV histogram models by learning, various images having different color characteristics that can be visually recognized, such as images 71 to 74 shown in FIG. You can also.

  Returning to FIG. 12, the down-sampling unit 52 performs a first process for reducing the resolution on the frame of interest, and compresses the resolution of each component of the HSV from 8 bits to 2-3 bits. After the second processing, the HSV component peak detection unit 53 and the attention area detection unit 55 are provided. The first process is a process performed mainly for improving the processing speed and removing noise in the image. The second process is a process performed mainly for the purpose of improving the robustness of the detection process by reducing the influence of illumination changes.

  The HSV component peak detection unit 53 calculates a 3D HSV histogram (such as graphs 71 to 73 in FIG. 13) for the frame of interest from the downsampling unit 52, and calculates H, S, and V from the 3D HSV histogram. Each component peak is detected, and each peak is notified to the region-of-interest range determination unit 54 as each central value (each average value) of H, S, and V.

  The region-of-interest range determination unit 54 stores the above-described H, S, and V stored as HSV histogram models in the HSV histogram model holding unit 51 for each central value (each peak) of the H, S, and V components. The respective ranges of H, S, and V having the respective dispersion widths are determined and notified to the attention area detection unit 55 as the attention area range.

  Specifically, for example, in the present embodiment, as described above, “16” as the range width of the H component, “128” as the range width of the S component, “64” as the range width of the V component, and HSV The histogram model holding unit 51 holds the HSV histogram model. Therefore, a range of H having a range width of 16 around the center value of the H component (H component peak), a range of S having a range width of 128 around the center value of the S component (peak of the S component), A range of V having a range width of 64 around the center value of the V component (peak of the V component) is determined as the attention area range and provided to the attention area detection unit 55.

  The attention area detection unit 55 sets each pixel constituting the attention frame from the down-sampling unit 52 as a pixel to be noticed as a target of sequential processing (hereinafter referred to as a attention pixel). By sequentially performing the processing, the attention area, that is, the play field area in the present embodiment is detected.

  That is, the attention area detection unit 55 determines whether or not the H, S, and V component values of the attention pixel are included in the attention area range notified from the attention area range determination unit 54, and determines that they are included. In such a case, the pixel of interest is detected as a pixel in the play field area, whereas when it is determined that the pixel of interest is not included, the pixel of interest is detected as a pixel other than the play field area.

  For example, in the example of FIG. 15, the image 82 is obtained as a result of detecting the play field region with the image 81 as the frame of interest. That is, the black area 91-1 in the image 82 is the detected play field area. Similarly, in the example of FIG. 16, the image 85 is obtained as a result of detecting the play field region with the image 84 as the frame of interest. That is, the black area 92-1 in the image 85 is the detected play field area.

  Thus, the attention frame in which the play field area is detected is provided from the attention area detection unit 55 in FIG. 12 to the post processing unit 56.

  The post processing unit 56 can remove a straight line or the like in the play field region by performing post processing, that is, filtering processing for noise removal, on the frame of interest. In other words, there is a process of re-determining a pixel that is determined to be a pixel in the play field area but is determined not to be a pixel in the play field area by the attention area detection unit 55 as a pixel in the play field area. It can also be understood that it is executed by the post-processing unit 56.

  For example, in the example of FIG. 15, an image 83 is obtained as a result of post-processing the image 82 described above. That is, the black area 91-2 in the image 83 is a play field area after white lines, players, and the like are removed from the area 91-1. Similarly, in the example of FIG. 16, an image 86 is obtained as a result of post-processing the image 85 described above. That is, the black area 92-2 in the image 86 is a play field area after white lines, players, and the like are removed from the area 92-1.

  As described above, the attention area extraction unit 11 in FIG. 12 has the color distribution characteristics obtained by learning various videos of the soccer game, that is, the HSV histogram model held in the HSV histogram model holding unit 51. Can be used to detect an image area having a turf color in a real-world play field of soccer, that is, a play field area in a robust and real-time manner. This makes it possible to automatically analyze soccer matches.

  In other words, as described above, the play field area is merely an example of the attention area when the video of the soccer game is adopted as the moving image content, and the image includes many attention colors according to the type of the moving image content employed. The region can be a region of interest. Therefore, when an attention area is detected from a predetermined type of video content, an HSV histogram model suitable for the predetermined type is generated by learning various videos of the predetermined type of video content, The HSV histogram model holding unit 51 may hold the HSV histogram model. Also in this case, if the predetermined type of moving image content is input to the attention area extraction unit 11, the attention area for the predetermined type is detected in a robust and real-time manner. This makes it possible to automatically analyze a predetermined type of moving image content.

  Although details will be described later, if an attention area is extracted, it is possible to easily extract object features in the attention area or in an area in contact with the attention area. Accordingly, it is possible to automatically provide meta information in analysis of various types of moving image content such as sports videos.

  Furthermore, as will be described in detail later, if object features can be extracted in a region of interest and its neighboring regions, highly accurate shot detection and shot classification in video analysis will be possible. Highlight extraction and automatic creation of digest video are also possible.

  The detailed example of the attention area extraction unit 11 in the image processing apparatus of FIG. 1 has been described above with reference to FIGS. 12 to 16.

  Next, a detailed example of the object feature extraction unit 12 in the image processing apparatus of FIG. 1 will be described with reference to FIGS. That is, hereinafter, some specific examples of object features that can be extracted by the object feature extraction unit 12 and the extraction method thereof when a soccer game is adopted as the moving image content will be described.

  FIG. 17 shows an example of an extraction method in the case where an object exists within the play field area of the frame of interest, and the size of the object, that is, the ratio of the object occupying the play field area is extracted as an object feature. FIG.

  Here, it is assumed that the pixel M (i, j) corresponding to the play field in the play field region of the frame of interest is expressed by Expression (1).

  The pixel M (i, j) corresponding to the play field is a pixel detected as a pixel in the play field area by the attention area detection unit 55 in FIG. A pixel before an object such as a player or a ball is removed. In other words, among the pixels constituting the play field area, the pixels excluding the pixels indicating objects such as players and balls existing in the play field area are the pixels M (I, j). That is, a pixel having the color characteristic of a play field (lawn) in the real world is a pixel M (I, j). Specifically, for example, the black pixels in the above-described play field region 91-1 in FIG. 15 and the play field region 92-1 in FIG. 16 correspond to the pixel M (I, J).

  In Expression (1), H and W indicate the sizes of the frame of interest in the vertical direction (Y direction) and the horizontal direction (X direction), respectively.

  In this case, the vertical direction projection feature amount Ph (i) and the horizontal direction projection feature amount Pv (j) shown in Expression (2) are used as feature amounts indicating the features of the object (hereinafter referred to as projection feature amounts). And each can be used.

  For example, FIG. 17 shows a graph 111-H representing the transition of the vertical projection feature amount Ph (i) and the transition of the horizontal projection feature amount Pv (j) when the image 101 is adopted as the frame of interest. A representing graph 111-V is shown.

  Details will be described later with reference to FIG. 32 and the like, but when the change in the horizontal direction projection feature amount Pv (j) is large and the change in the vertical direction projection feature amount Ph (i) is small, a large object is present in the play field area. It means that there is an object such as a player. On the other hand, when the change in the horizontal projection feature amount Pv (j) is small, or when the change in the vertical projection feature amount Ph (i) is large, there is a small object inside the play field area, for example, a player or the like. This means that the object is small.

  As described above, the object feature extraction unit 12 calculates the vertical direction projection feature amount Ph (i) and the horizontal direction projection feature amount Pv (j) for the frame of interest, and calculates the vertical direction projection feature amount Ph (i) and the horizontal projection feature amount Ph (i). Based on the degree of each change with the direction projection feature amount Pv (j), the size of the object existing inside the play field region can be extracted as one of the object features.

  FIG. 18 shows an extraction method in the case where the boundary between the play field region of the frame of interest and other regions, or information about the boundary, for example, the angle of the boundary line and the intersection position with the other boundary line are extracted as object features. It is a figure explaining an example.

  That is, FIG. 18 shows an example in which straight lines 123 and 124 are extracted as the boundaries of the black play field region 122 when the image 121 is adopted as the frame of interest.

  The object feature extraction unit 12 detects the position of the boundary of the play field area of the frame of interest in units of pixels. Hereinafter, the position of the boundary detected in this way is described as B (i) as shown in Equation (3).

  In this case, since each position B (i) of the boundary indicates each position for W pixels in the horizontal direction, the object feature extraction unit 12 performs a resolution reduction process to divide into 30 in the horizontal direction, that is, The number of horizontal positions B (i) of the boundary is reduced to 30 and each angle A (i) of the boundary line at each position B (i) of the boundary, for example, each position B (i) And each angle A (i) of a straight line connecting the positions B (i−1), B (i + 1), and the like before and after that. The purpose of lowering the resolution is to reduce the influence (noise) of the person or graphic that touches the boundary, that is, the person or graphic that is detected as the boundary position B (i) is not an actual boundary. This is to remove the position of the device.

  After that, the object feature extraction unit 12 calculates the angle Al (n) of the left side n (n is an integer value between 1 and 30) of 30 divided in the horizontal direction according to the equation (4), The right angle Ar (n) is calculated according to the equation (5).

  Further, the object feature extraction unit 12 calculates each of the variance values Dl (n) and Dr (n) of the left and right angles according to the equations (6) and (7).

Then, the object feature extraction unit 12 sets the pixel position N represented by Expression (8) as the intersection position of the left and right straight lines.

  Finally, the object feature extraction unit 12 determines that the angle difference between the left and right angles Al (n) and Ar (n) at the pixel position N is 10 degrees or more (that is, in the case of Expression (9)). Each straight line having left and right angles Al (n) and Ar (n) is detected as two different boundary lines. For example, in the example of FIG. 18, the straight lines 123 and 124 are detected as two boundary lines. If the angle difference between the left and right angles Al (n) and Ar (n) at the pixel position N is less than 10 degrees, it is regarded as one boundary line.

  When two boundary lines are detected, as shown in FIG. 21 described later, the intersection position of the two boundary lines can be extracted as a corner position in the soccer field.

  In this way, the object feature extraction unit 12 uses the boundary between the play field area of the frame of interest and another area, or information about the boundary, for example, the angle of the boundary line or the intersection position with the other boundary line as the object feature. Can be extracted.

  Furthermore, after detecting the boundary of the play field area, the object feature extraction unit 12 uses a goal (hereinafter referred to as a goal gate, goal box, etc.) existing in the soccer field as one of the object features as follows. It can also be extracted.

  That is, the object feature extraction unit 12 detects a white line in a substantially vertical direction in contact with each boundary line of the frame of interest. Here, the term “substantially vertical” means that the vertical part of the white structure (rebar, etc.) of the goal gate in the real world is not always accurate when it is included as a white line image in the frame of interest. This is because it is not vertical in the frame. In other words, the object feature extraction unit 12 is a white line that touches each boundary line and has an angle in the grace range of α degrees to the left and right (α is an arbitrary value from 0 to 90 degrees) with respect to the vertical direction in the frame of interest. Is detected. Further, the white line here does not mean a true white line, but means a line composed of colors within a range that can be regarded as white.

  Then, when two or more white lines in a substantially vertical direction are detected, the object feature extraction unit 12 selects a predetermined two sets of the two or more white lines, and uses the two white lines of the set. It is confirmed whether or not a white line in a substantially horizontal direction exists in the surrounding area. Here, the substantially horizontal direction is the same as the above-described substantially vertical direction. Such confirmation is performed for all combinations that can be combined. When the object feature extraction unit 12 detects one or more sets of white lines in the substantially horizontal direction among the sets of two white lines in the substantially vertical direction, the object feature extraction unit 12 treats these sets as candidate goal gates.

  Then, the object feature extraction unit 12 has dimensions such as the relationship between the post height 133 of the goal gate 131 and the crossbar length 132 shown in FIG. 19 and the proportional relationship between the crossbar length 132 and the play field width. The correct goal gate is extracted using the rule, that is, the semantic rule of the goal gate. That is, the object feature extraction unit 12 selects a candidate satisfying the semantic rule from among one or more goal gate candidates as a goal gate.

  FIG. 20 shows a specific example of the goal gate detected by such a method. That is, in FIG. 20, when the image 141 is adopted as the frame of interest, two substantially vertical white line segments in contact with the boundary line of the black play field region 143 and one substantially horizontal line therebetween. An example in which a region 144 surrounded by white line segments in the direction is extracted as a goal gate is shown. FIG. 20 shows that the goal gate 144 can be detected robustly even though the image 141 has a complicated background.

  FIG. 21 is a diagram for explaining an example of the extraction method when the corner position in the play field area of the frame of interest is extracted as an object feature.

  That is, FIG. 21 shows an example in which the intersection position 164 of the two boundary lines 162 and 163 of the black play field region 161 is extracted as the corner position when the image 151 is adopted as the frame of interest. ing.

  The corner position detection method is basically the same as the boundary line detection method described above with reference to FIG. That is, the object feature extraction unit 12 can extract the corner position as one of the object features by calculating the above-described equations (3) to (9) for the frame of interest.

  In addition, a specific example of the extraction method is redundant and will not be described. However, the object feature extraction unit 12 can extract various object features by using a known method or the like.

  Examples of specific object feature extraction results are shown in FIGS.

  That is, in the example of FIG. 22, the player 171 existing in the play field area is extracted as one of the object features. In the example of FIG. 23, a graphic area 172 such as a score display synthesized on the play field area is extracted as one of the object features. In the example of FIG. 24, the ball 173 existing in the play field area is extracted as one of the object features. In the example of FIG. 25, for example, a judge 174, not a player, is extracted as one of the object features using a difference in clothes color between the player and the judge. In the example of FIG. 26, the movement of the player existing in the play field area, that is, the movement of the player running at a predetermined speed in a predetermined direction is extracted as one of the object features. In the example of FIG. 27, the player's face 175 is extracted as one of the object features.

  The detailed example of the object feature extraction unit 12 in the image processing apparatus of FIG. 1 has been described above with reference to FIGS. 17 to 27.

  Next, a detailed example of the shot cut detection unit 13 in the image processing apparatus of FIG. 1 will be described with reference to FIGS.

  Shot cut refers to switching between shots, that is, shot change, and is generally roughly divided into three types: hard cut, fade-in / fade-out cut, and graphics wiper cut. With hard cuts, shots are switched between 1 and 2 frames, so that shot changes are easy to detect, but with fade-in / fade-out cuts and graphics wiper cuts, shots are switched between 10 and 30 frames. Therefore, it is not easy to automatically detect the shot change.

  Therefore, the present inventor has invented the following shot cut detection method in order to easily and accurately detect not only hard cut but also fade-in / fade-out cut and graphics wiper cut.

  That is, the present inventor considered that the easy and accurate detection of shot cuts is based on the first condition of maintaining robustness against camera movement and object movement within the frame, and changes in the color feature distribution within the frame. It is necessary to satisfy the second condition that it can be detected sensitively at the same time.

  In this case, in order to satisfy the first condition and the second condition, the following various operations are performed by dividing the image into two or more image areas with respect to the dominant direction of movement of the object in the frame. The present inventor thought.

  As the dominant direction of the movement of the object in the frame, the horizontal direction of the frame can be adopted when the moving image content is a video of a soccer game as in the present embodiment. That is, soccer game objects are players, balls, and the like, and these players and balls often move in a horizontal direction with respect to the frame.

  Specifically, for example, in the present embodiment, as shown in FIG. 28, for each of the frames 181 to 183, the image area is horizontally divided into three at a ratio of 2: 5: 3 from the top. The following various operations are performed. Hereinafter, in accordance with the description of FIG. 28, each of the image areas divided into 2: 5: 3 is referred to as areas R1, R2, and R3.

  Here, the reason for the division into 3: 5: 3 is the analysis of the various frames that make up one scene of a soccer game. Because it becomes. That is, it is not always necessary to adopt the 3: 5: 3 division as the number of divisions and the division ratio, and it can be arbitrarily changed according to the characteristics of the type of moving image content for which shot cut is desired to be detected.

  As described above, the shot cut detection unit 13 first divides the frame of interest into regions R1 to R3.

  Next, the shot cut detection unit 13 calculates the 3D HSV histogram (see graphs 71 to 73 in FIG. 13) described above as the color feature distribution for each of the regions R1 to R3 of the frame of interest. In addition, the shot cut detection unit 13 represents each H, S, V component of the 3D HSV histogram as 2 bits in order to improve robustness against noise in the frame and the like. The 6-bit data is generated as a feature quantity indicating a color feature distribution. Hereinafter, such a feature amount is described as a color feature amount HRk. Here, k represents any one of 1 to 3. That is, the color feature value HR1 is a color feature value for the region R1, the color feature value HR2 is a color feature value for the region R2, and the color feature value HR3 is a color feature value for the region R3. It is.

  Then, the shot cut detection unit 13 calculates a feature amount ΔHk of the following equation (10) for each of the regions R1 to R3 of the frame of interest as one of the features indicating the difference between frames. Hereinafter, the feature amount ΔHk is referred to as a color difference feature amount ΔHk.

  In Expression (10), Htk (i) represents the color feature amount for the region k of the frame of interest. Htk-Δ (i) indicates the color feature amount for the region k of the frame several frames before the target frame. That is, Δ is a constant indicating how many frames ago the difference from the feature value is obtained.

  As a result, the color difference feature amount ΔHR1 for the region R1, the color difference feature amount ΔHR2 for the region R2, and the color difference feature amount ΔHR3 for the region R3 are obtained.

  Therefore, the shot cut detection unit 13 calculates the color difference feature amount ΔH for the entire frame by calculating the following equation (11). Hereinafter, the color difference feature amount ΔH is referred to as a total color difference feature amount ΔH.

  In equation (11), α, β, and γ are respectively the color difference feature amount ΔHR1 for the region R1, the color difference feature amount ΔHR2 for the region R2, and the color difference feature amount ΔHR3 for the region R3. It is a predetermined parameter value for giving weight to each. Here, the specific values of α, β, and γ are not particularly limited, but it is preferable to set β to the largest value for a soccer game as in the present embodiment. As described above, since the area R2 includes the most play field areas, it is the most important area for shot cut detection, and it is appropriate to give a large weight to the area R2.

  Then, the shot cut detection unit 13 compares the total color difference feature amount ΔH with a preset threshold value T1, and when the total color difference feature amount ΔH is larger than the threshold value T1 (ΔH> T1 ), The target frame is detected as a shot cut. The threshold T1 is not particularly limited, but it is preferable to adopt a value that is statistically obtained from a video such as a soccer game.

  A technique corresponding to a series of processes executed by the above shot cut detection unit 13 is one of the shot cut detection techniques invented by the present inventors. Hereinafter, this method is referred to as a comprehensive color feature difference shot cut detection method.

  Furthermore, the present inventors have also invented the following shot cut detection method.

  That is, in a sport video such as a soccer game, it is possible to detect shot cuts more robustly by using the feature of the attention area. That is, the attention area of the soccer game is a play field area, and the area in the frame including the most play field areas is the above-described area R2.

  Therefore, shot cuts can also be detected using the feature of the target color in the region R2. Here, the attention color is a dominant color in the attention area, and when the attention area is the play field area, it is the color of the turf, that is, so-called green. In other words, the attention color area corresponds to the play field area before post-processing by the post-processing unit 56 of FIG. 12, that is, the grass in the real world that does not include objects such as players, balls and white lines. It is an area to do.

  Specifically, the shot cut detection unit 13 calculates the ratio of the target color area in the region R2, that is, the ratio of the number of target color pixels to the total number of pixels constituting the region R2. Hereinafter, this ratio is referred to as a noticeable color rate DCR.

  Next, the shot cut detection unit 13 determines whether or not the target color rate DCR exceeds 50%. If it is determined that the target color rate DCR exceeds the target color rate DCR, it is determined that the target frame may be shot cut, and the following formula ( The calculation of 12) is performed. That is, Expression (12) is also an expression for calculating a feature amount ΔDCR that is one of the features indicating the difference between frames. Hereinafter, the feature amount ΔDCR is referred to as a noticeable color difference feature amount ΔDCR.

  In the equation (12), DCRt (i) represents the target color rate for the target frame. DCRt-Δ indicates the target color rate of a frame several frames before the target frame. That is, Δ is a constant indicating how many frames ago the difference from the feature amount is obtained, as in the above-described equation (10).

  Then, the shot cut detection unit 13 compares the target color difference feature amount ΔDCR with a preset threshold value T2, and when the target color difference feature amount ΔDCR is larger than the threshold value T2 (ΔDCR> T2 ), The target frame is detected as a shot cut. The threshold T2 is not particularly limited, but it is preferable to adopt a value that is statistically obtained from a video such as a soccer game.

  A technique corresponding to a series of processes executed by the above shot cut detection unit 13 is one of the shot cut detection techniques invented by the present inventors. Hereinafter, this method is referred to as a noticed color difference shot cut detection method.

  Note that the total color feature difference shot cut detection method and the target color difference shot cut detection method are independent methods, and therefore it is possible to detect a shot cut using only one of them. However, by combining the total color feature difference shot cut detection method and the target color difference shot cut detection method, a more accurate shot cut can be detected.

  For example, FIG. 29 shows an example of shot cut detection processing when the total color feature difference shot cut detection method and the target color difference shot cut detection method are applied in combination, that is, an example of detailed processing in step S3 of FIG. A flowchart to explain is shown.

  In step S21, the shot cut detection unit 13 determines the input frame as the frame of interest.

  In step S22, the shot cut detection unit 13 calculates the color difference feature amount ΔH for the frame of interest by the calculation of the above-described equation (11) or the like.

  In step S22, the shot cut detection unit 13 determines whether or not the color difference feature amount ΔH is larger than the threshold value T1 (whether ΔH> T1).

  When the color difference feature amount ΔH exceeds the threshold value T1, it is determined as YES in the process of step S23, and the process proceeds to step S24. In step S24, the shot cut detector 13 detects that the frame of interest is a shot cut. Thereby, a process progresses to step S30. However, the processing after step S30 will be described later.

  On the other hand, when the color difference feature amount ΔH is equal to or smaller than the threshold value T1, it is determined NO in the process of step S23, and the process proceeds to step S25. In step S <b> 25, the shot cut detection unit 13 calculates a target color rate DCR for the target frame.

  In step S <b> 26, the shot cut detection unit 13 determines whether the target color rate DCR exceeds 50% (whether DCR> 50%).

  If the target color rate DCR is 50% or less, it is determined NO in the process of step S26, and the process proceeds to step S27. In step S27, the shot cut detector 13 detects that the frame of interest is not a shot cut. Thereby, a process progresses to step S30. However, the processing after step S30 will be described later.

  On the other hand, if the target color rate DCR exceeds 50%, it is determined YES in the process of step S26, and the process proceeds to step S28. In step S <b> 28, the shot cut detection unit 13 calculates the target color difference feature amount ΔDCR by the calculation of the above-described equation (12) or the like.

  In step S29, the shot cut detection unit 13 determines whether or not the target color difference feature amount ΔDCR is larger than the threshold value T2 (whether ΔDCR> T2).

  If the target color difference feature amount ΔDCR is larger than the threshold value T2, it is determined YES in the process of step S29, and the target frame is detected as a shot cut in the process of step S24.

  On the other hand, when the target color difference feature amount ΔDCR is equal to or smaller than the threshold T2, it is determined NO in the process of step S29, and it is detected that the target frame is not a shot cut in the process of step S27. The

  In this way, when the process of step S24 or S27 ends, that is, when the detection of whether or not the frame of interest is a shot cut ends, the process proceeds to step S30. In step S30, the shot cut detection unit 13 determines whether or not the frame of interest is the last frame.

  If it is determined in step S30 that the frame of interest is not the last frame, the process returns to step S21, and the subsequent processes are repeated.

  That is, the loop processing of steps S21 to S30 described above is repeated for each frame constituting the moving image content, and detection of whether or not it is a shot cut is performed for each frame. When it is detected whether or not the last frame is a shot cut, it is determined YES in the process of step S30, and the shot cut detection process ends. That is, the process of step S3 in FIG. 2 ends, and the shot classification process of step S4 is continued.

  The detailed example of the shot cut detection unit 13 in the image processing apparatus of FIG. 1 has been described above with reference to FIGS. 28 and 29. By adopting such a shot cut detection unit 13, the entire image (frame) is divided into a plurality of regions that can be described with feature quantities that are robust to the horizontal motion of the camera and the horizontal motion of the person, respectively. By using the color distribution feature and the region feature of the target color, it is possible to detect shot cuts with high accuracy, and as a result, it is possible to automatically add meta information in the analysis of sports images and the like.

  Next, a detailed example of the shot classification unit 14 in the image processing apparatus of FIG. 1 will be described with reference to FIGS. 30 to 36.

  That is, as a result of the shot cut detection process of the shot cut detection unit 13 described above, the moving image content is divided into a plurality of shots. Therefore, the shot classification unit 14 classifies each of the plurality of shots one by one into a predetermined type of one or more types defined in advance. Such a type is also called a shot type as will be described later.

  In the present embodiment, for example, “Long”, “Medium”, “Close-up”, “Close-up”, “as shown in FIG. Assume that four types of “Out” are defined.

  As can be seen from each sample image in FIG. 30, “Long” refers to a type of shot that shows the overall situation of a soccer game, that is, a shot composed of an overall view of a play field area and players. “Medium” is a type of shot in which several players and the like are present in the play field area with the camera zoomed. “Close-up” refers to a type of shot in which one or two players are focused and their faces can be identified. “Out” refers to a type of shot that does not include a play field area.

  FIG. 31 shows a flowchart for explaining a detailed example of the shot classification process classified into one of the four types, that is, the process of step S4 in FIG.

  In step S51, the shot classification unit 14 sets the input shot as a target shot.

  In step S52, the shot classification unit 14 determines whether or not the attention shot includes the attention area, that is, in the present embodiment, whether or not the play field area is included.

  If it is determined in step S52 that the attention area does not include the attention area, the shot classification unit 14 determines in step S53 that the attention shot type is “Out” or “Close-up” outside the field. To do. Thereafter, the process proceeds to step S62. However, the processing after step S62 will be described later.

  Note that “Close-up” outside the field refers to a type of shot in which one or two players or the like are focused and whose face can be identified, and in which the play field area is almost not detected. In this sense, “Close-up” outside the field can be included in “Out”, but it is also possible to use a face identified from “Close-up” outside the field in highlight detection described later. Therefore, it is described here separately from “Out”.

  On the other hand, when the play field area is included in the target shot, it is determined as YES in the process of step S52, and the process proceeds to step S54.

  In step S54, the shot classification unit 14 calculates the projection feature amounts Ph and Pv for the attention area of the attention shot.

  Specifically, for example, the shot classification unit 14 uses the above-described formula (2) to calculate each vertical projection feature amount Ph (i) for H pixels in the vertical direction and each W pixel in the horizontal direction. A horizontal projection feature amount Pv (j) is calculated.

  Further, in order to reduce the influence of noise, the shot classification unit 14 performs low resolution processing on each vertical projection feature amount Ph (i) for H pixels in the vertical direction, and each vertical direction for 10 pixels. The projection feature amount Ph (i) is also applied to each horizontal projection feature amount Pv (j) for W pixels in the horizontal direction, and low-resolution processing is performed, so that 30 horizontal projection feature amounts Pv (j ).

  As a result of such processing, as each of the projection feature amounts Ph and Pv for the attention area of the attention shot, a feature amount indicating the distribution of each of the vertical projection feature amounts Ph (i) for ten pieces, and 30 Thus, a feature amount indicating a distribution of each horizontal projection feature amount Pv (j) is obtained.

  That is, the feature quantity indicating the distribution of each vertical projection feature quantity Ph (i) as shown in the graph 111-H in FIG. 17 described above is an example of the vertical projection feature quantity Ph. In other words, when a pixel having a color characteristic of a play field (a pixel excluding an object) among pixels constituting the play field region is projected onto one line in the vertical direction, a feature amount indicating a distribution on the one line is This is the vertical projection feature amount Ph.

  Also, the feature quantity indicating the distribution of each horizontal projection feature quantity Pv (j) as shown in the graph 111-V of FIG. 17 described above is an example of the horizontal projection feature quantity Pv. In other words, when a pixel having a color characteristic of a play field (a pixel excluding an object) among the pixels constituting the play field region is projected onto one line in the horizontal direction, the feature amount indicating the distribution on the one line is This is the horizontal projection feature amount Pv.

  In step S55, the shot classification unit 14 determines whether or not the feature based on the projected feature amounts Ph and Pv for the attention area of the attention shot is of the “Close-up” or “Medium” type.

  The features based on the projection feature amounts Ph and Pv for the attention area of the attention shot are not particularly limited as long as the types of shots can be classified, but here, the features shown in the table of FIG. 32 are adopted. Suppose that

  The “Close-up” or “Medium” type means that the type of the shot of interest is classified as “Close-up” or “Medium”. The other types described later are basically the same.

  That is, FIG. 32 shows the relationship between the projection feature amounts Ph and Pv for the shot attention area and the shot type.

  The Vertical Projection Feature on the horizontal axis in FIG. 32 is the value of 30 horizontal projection feature amounts Pv (j) obtained as a result of the same processing as the processing in step S54 described above for a predetermined shot. The difference value between the maximum value and the minimum value is shown. That is, the difference value between the maximum value and the minimum value of the distribution indicated by the horizontal direction projection feature amount Pv is the Vertical Projection Feature.

  On the other hand, the horizontal projection feature on the vertical axis in FIG. 32 is the vertical projection feature amount Ph (i) for 10 obtained as a result of the same processing as the processing in step S54 described above for a predetermined shot. The difference value between the maximum value and the minimum value is shown. That is, the difference value between the maximum value and the minimum value of the distribution indicated by the vertical direction projection feature amount Ph is the horizontal projection feature.

  Further, in FIG. 32, a small dot marked with * is obtained by performing a process similar to the process in step S54 described above on one shot known to be “Long” and plotting the process result. Show. A small dot marked with ○ indicates that one shot known to be “Medium” is subjected to the same processing as the processing in step S54 described above, and the processing result is plotted. The small □ marks indicate the result of performing the same processing as the processing in step S54 described above on one shot known to be “Close-up” and plotting the processing result.

  Note that plotting the result of the process similar to the process of step S54 described above on the table of FIG. 32 uses the projection feature amounts Pv and Ph obtained as a result of the process similar to the process of step S54, respectively. Specifically, the maximum value of the distribution indicated by the vertical direction projection feature amount Ph, with the difference value between the maximum value and the minimum value of the distribution indicated by the horizontal direction projection feature amount Pv as the coordinate of the horizontal axis. The point is plotted on the table of FIG. 32 using the difference value between the minimum value and the coordinate of the vertical axis.

  As a result of each plot for a plurality of shots of which the type is known, the region 191 includes “Long” or “Medium” shot points. Thereby, when the result of the process of step S54 described above is plotted on the table of FIG. 32, when the plotted point is included in the area 191, the shot of interest can be classified as “Long” or “Medium”. Recognize.

  Note that the plotting of the result of the process of step S54 described above on the table of FIG. 32 means plotting using the projection feature amounts Pv and Ph obtained as a result of the process of step S54 described above, specifically, The difference value between the maximum value and minimum value of the distribution indicated by the horizontal direction projection feature amount Pv is the horizontal axis coordinate, and the difference value between the maximum value and minimum value of the distribution indicated by the vertical direction projection feature amount Ph is Plotting points on the table of FIG. 32 as axis coordinates.

  Similarly, as a result of each plot for a plurality of shots whose types are known, the area 192 includes “Long” shot points. Thereby, when the result of the process of step S54 described above is plotted on the table of FIG. 32, when the plotted point is included in the region 192, it is understood that the shot of interest can be classified as “Long”.

  As a result of each plot for a plurality of shots whose types are known, the region 193 includes all shot points of “Long”, “Medium”, and “Close-up”. In other words, when the result of the processing in step S54 described above is plotted on the table of FIG. 32, when the plotted point is included in the region 193, it can be understood that it is difficult to classify the shot of interest. In other words, it can be said that the region 193 is a so-called gray zone.

  As a result of each plot for a plurality of shots whose types are known, the region 194 includes “Close-up” or “Medium” shot points. Thereby, when the result of the process in step S54 described above is plotted on the table of FIG. 32, if the plotted point is included in the region 194, the shot of interest is classified as “Close-up” or “Medium”. I understand that I can do it.

  Although the region 193 is a gray zone that is difficult to classify as “Long”, “Medium”, or “Close-up”, it is actually classified as “Long”. It can be said that the region is rare and close to the region 194. Therefore, here, in order to simplify the classification of the types of shots, when the result of the processing in step S54 described above is plotted on the table of FIG. 32, the plotted point is included in the region 193 or the region 194. In some cases, the shot of interest is classified as “Close-up” or “Medium”.

  That is, when the result of the process of step S54 described above is plotted on the table of FIG. 32, if the plotted point is included in the region 193 or the region 194, it is determined as YES in the process of step S55. That is, it is determined that the feature based on the projected feature amounts Ph and Pv for the attention area of the attention shot is of the “Close-up” or “Medium” type, and the process proceeds to step S56.

  In step S56, the shot classification unit 14 determines whether or not the object feature is of the “Close-up” type.

  If it is determined in step S56 that the object feature is of the “Close-up” type, the shot classification unit 14 determines in step S57 that the type of shot of interest is “Close-up”. Thereafter, the process proceeds to step S62. However, the processing after step S62 will be described later.

  On the other hand, when it is determined in step S56 that the object feature is not “Close-up” type, that is, when it is determined that the object feature is “Medium” type, the shot classification unit 14 in step S58. The type of shot of interest is determined to be “Medium”. Thereafter, the process proceeds to step S62. However, the processing after step S62 will be described later.

  That is, the case where it is determined as YES in the process of step S55 is a case where the shot of interest can be classified as “Close-up” or “Medium”.

  In such a case, in this embodiment, the shot of interest is classified as “Close-up” or “Medium” based on the object feature extracted by the object feature extraction unit 12 of FIG. 1 described above. I did it.

  In this case, there are various types of object features extracted by the object feature extraction unit 12, any type can be used, and one or more types can be used in combination.

  For example, in the example of FIG. 33, a face area is used as an object feature. That is, the face areas 211 and 212 are detected from the image 201, while the face area 213 is detected from the image 202. When comparing the face areas 211 and 212 with the face area 213, it can be seen that the sizes are different. Therefore, the size of the face area is defined as a feature value, a type whose feature value exceeds the threshold value is defined as a “Close-up” type, and a type whose feature value is equal to or less than the threshold value is defined as a “Medium” type. can do. That is, the shot classifying unit 14 determines “Close-Up” for a shot of interest that includes a face area having a size exceeding the threshold, while “shot-up” includes “ It can be determined to be “Medium”.

  Even if the face area is not extracted by the object feature extraction unit 12, for example, if the player's area is extracted using the color of the player's uniform, the shot classification unit 14 “Medium” or “Close-up” can be classified using the width of the feature as a feature amount.

  In this way, when the result of the process of step S54 described above is plotted on the table of FIG. 32, if the plotted point is included in the region 193 or the region 194, YES in the process of step S55. As a result of the determination, the type of the shot of interest is determined to be “Close-Up” or “Medium” by the processing of steps S56 to S58.

  On the other hand, when the result of the process of step S54 described above is plotted on the table of FIG. 32, if the plotted point is included in the area 191 or the area 192, NO in the process of step S55. As a result of the determination, the following processing from step S59 onward is executed, and the type of shot of interest is determined to be “Medium” or “Long”.

  That is, in step S59, the shot classification unit 14 calculates the projection feature amounts Ph1 and Pv1 for the background excluded region in the attention region of the attention shot.

  Here, the background exclusion area is an area excluding the background area indicating the spectator seat portion and the like in the image of the soccer game. Specifically, for example, in the example of FIG. 34, the soccer game image 231 is a background area other than the play field area 241. Therefore, the background excluded area in the soccer game image 231 is the area below the solid line 251 and between the dotted lines 252 and 253 is used as the background excluded area. Similarly, in the soccer game image 232, the area other than the play field area 242 is the background area. Therefore, the background exclusion area in the soccer game image 232 is below the solid line 254 and the area between the dotted lines 255 and 256 is used as the background exclusion area.

  Therefore, the shot classification unit 14 uses the above-described equation (2) for such a background exclusion region, and uses the vertical projection feature amount Ph (i) for H pixels in the vertical direction and the W in the horizontal direction. Each horizontal projection feature amount Pv (j) for the pixel is calculated.

  Furthermore, in order to reduce the influence of noise, the shot classification unit 14 performs low resolution processing on the vertical projection feature amount Ph (i) for H pixels in the vertical direction, and projects each of 10 vertical projections. A low-resolution process is performed on the horizontal projection feature amount Pv (j) for W pixels in the horizontal direction as the feature amount Ph1 (i), and 30 horizontal projection feature amounts Pv1 (j) To do.

  As a result of such processing, as the projection feature amount Ph1, Pv1 for the background exclusion region, the feature amount indicating the distribution of each of the 10 vertical projection feature amounts Ph1 (i), and each of the 30 feature amounts A feature amount indicating the distribution of the horizontal direction projected feature amount Pv1 (j) is obtained.

  Therefore, in step S60, the shot classification unit 14 determines whether or not the feature based on the projection feature amounts Ph1 and Pv1 for the background excluded region is a “Medium” type.

  That is, when each of the plurality of shots plotted in the area 191 of FIG. 32 is re-plotted using each background exclusion area, the “Long” background exclusion area points are as shown in FIG. While included in the area 271, the point of the background exclusion area of “Medium” is included in the area 272 of the area 191.

  Thus, when the processing result of step S59 described above is plotted on the tables of FIGS. 32 and 36, when the plotted point is included in the region 272, it is determined as YES in the processing of step S60. That is, it is determined that the feature based on the projection feature amounts Ph1 and Pv1 for the background excluded region is the “Medium” type, and in Step S58, the type of the shot of interest is determined to be “Medium”.

  On the other hand, when the result of the process of step S59 described above is plotted on the tables of FIGS. 32 and 36, the plotted point is included in the region 192 (FIG. 32) or the region 271 (FIG. 36). Sometimes, it is determined as NO in the process of step S60, that is, it is determined that the feature based on the projection feature amounts Ph1 and Pv1 for the background excluded region is the “Long” type, and the process proceeds to step S61. In step S61, the shot classification unit 14 determines that the type of the shot of interest is “Long”.

  As described above, when the processing of step S53, S57, S58, or S61 is completed, that is, the type of the shot of interest is “Out”, “Out-field” “Close-up”, “Close-up”, “Medium”. Or “Long”, the process proceeds to step S62.

  In step S62, the shot classification unit 14 determines whether or not the target shot is the last shot.

  If it is determined in step S62 that the target shot is not the last shot, the process returns to step S51, and the subsequent processes are repeated.

  That is, for each shot in the moving image content, the loop processing of steps S51 to S62 described above is repeated, and the type of each shot is determined. When the type of the last shot is determined, it is determined as YES in the process of step S62, and the shot classification process ends. That is, the process of step S4 in FIG. 2 ends, and the highlight detection process of step S5 is continued.

  The detailed example of the shot classification unit 14 in the image processing apparatus of FIG. 1 has been described above with reference to FIGS. 30 to 36. By adopting such a shot classification unit 14, for various shot types, projection features in the attention area (for example, the above-described projection feature amounts Ph and PV) and objects in the attention area and adjacent areas Features can be used, resulting in highly accurate shot classification

  Next, a detailed example of the highlight detection unit 15 in the image processing apparatus of FIG. 1 will be described with reference to FIGS.

  FIG. 37 shows a framework for highlight extraction by the highlight detection unit 15. As shown in FIG. 37, highlight extraction is basically performed in three layers Layer1 to Layer3.

  Layer 1 is a layer representing a shot type temporal relevance rule between shots or a statistical model for such a relevance rule. Here, the shot type refers to the type of each shot classified by the shot classification unit 14. For example, in the present embodiment for a soccer game, “Long”, “Medium”, There are four types, “Close-up” and “Out”.

  For example, a plurality of shot groups that are known to be soccer corner kick scenes to be extracted as highlights are prepared as learning images, and rules for time transitions of shot types are learned from the plurality of learning images, If there is a statistical model indicating the rules obtained as a result of the learning in advance, the highlight detection unit 14 recognizes the shot type temporal transition of the newly input moving image content, for example, as shown in FIG. If you find a shot group whose shot type time transition matches or is similar to the statistical model, highlight that shot group as a highlight. Can do.

  Layer 2 is a layer representing a temporal relevance rule between frames having a meaning among frames constituting a predetermined shot, or a statistical model for such a relevance rule.

  For example, a plurality of shots that are known to be soccer corner kick scenes to be extracted as highlights are prepared as learning images, and the frame type of each frame constituting one shot is prepared from the plurality of learning images. If the rule of time transition is learned and a statistical model indicating the rule obtained as a result of the learning exists in advance, the highlight detection unit 14 recognizes the time transition of the frame type in each shot for the moving image content. For example, if a frame type time transition as shown in FIG. 10 is recognized and a shot that matches or resembles a statistical model is found, the shot can be extracted as at least a part of the highlight. . The frame type such as “Upper Corner” shown in FIG. 10 will be described later with reference to FIG.

  Layer 3 is a layer that assigns meaning to the frame based on the feature amount in the frame. That is, a layer obtained by using a frame defined by Layer 3, that is, a layer representing a temporal relevance rule between frames is Layer 2. Here, as the feature amount in the frame, for example, in this embodiment, the feature amount related to the soccer game, for example, the various object features described above with reference to FIGS. 17 to 27, that is, the object feature extraction unit 12 in FIG. Various extracted object features can be employed. Alternatively, when the frame type transition is used as in the above-described example as Layer 2, the frame type can be regarded as a feature amount in the frame. The frame type can be determined based on the object feature extracted by the object feature extraction unit 12 of FIG. 1, and this will be described later with reference to FIG. 42.

  FIG. 38 shows a detailed functional configuration example of the highlight detection unit 15.

  The highlight detection unit 15 in the example of FIG. 38 is configured to include a learning unit 301 to a highlight extraction unit 306.

  The learning unit 301 learns the rules for Layer 1 and Layer 2 described above, and causes the model holding unit 302 to hold a statistical model obtained as a result of the learning. Details of the learning unit 301 will be described later with reference to FIGS. 40 to 44.

  The moving image content from the shot type classification unit 14 in FIG. 1 is input to the Layer 1 feature amount generation unit 303 and the Layer 3 feature amount generation unit 304. That is, the moving image content input to the Layer1 feature value generation unit 303 and the Layer3 feature value generation unit 304 is obtained by extracting the object features for each frame by the attention area extraction unit 11 and the object feature extraction unit 12 in FIG. The moving image content is divided into a plurality of shots by the cut detection unit 13, and each shot type of the plurality of shots is determined by the shot classification unit 14.

  Therefore, the Layer1 feature amount generation unit 303 generates a feature amount (hereinafter referred to as Layer1 feature amount) for performing highlight extraction in Layer1, and provides the feature amount to the highlight extraction unit 306.

  Note that various feature quantities can be adopted as the Layer 1 feature quantity in accordance with the form of the Layer 1 statistical model held in the model holding unit 302.

  Specifically, for example, when a shot type time transition order or a state transition diagram with each shot type as each state is given as a statistical model of Layer 1, a shot as shown in FIG. Type transition over time can be used as Layer1 feature.

  Also, for example, as a statistical model of Layer1, there is a total period (length) of “2” non- “Long” shots where “1) the number of non-“ Long ”shots before“ Long ”shots is within a predetermined range” When a condition such as a length within a predetermined range is satisfied, a soccer corner kick scene is given. If a model is given, specify whether the conditions 1) and 2) are satisfied. Information that can be obtained, that is, information such as the number of non- “Long” shots before a “Long” shot to be noted as a processing target (hereinafter referred to as “Long” shot of interest ”), and the total period (length) of the non-Long shot Can be adopted as the Layer1 feature amount.

  As described above, the Layer3 feature value generation unit 304 generates information such as the object feature and frame type in the frame as the Layer3 feature value and provides it to the Layer2 feature value generation unit 305.

  The Layer2 feature value generation unit 305 generates a feature value (hereinafter referred to as a Layer2 feature value) for performing highlight extraction in Layer2, and provides the feature value to the highlight extraction unit 306.

  Note that various feature quantities can be adopted as the Layer 2 feature quantity according to the form of the Layer 2 statistical model held in the model holding unit 302.

  Specifically, for example, when the order of transition of frame types as a statistical model of Layer 2 or a state transition diagram as shown in FIG. 44 to be described later is given, the frame type as shown in FIG. This time transition can be used as the Layer 2 feature amount.

  In addition, for example, as a statistical model of Layer 2, “1) a goal box is observed over several frames, 2) no center line is observed in front of the goal box, and 3) the goal box is centered from both sides of the screen. When the model is given, “If the frame moves so that it moves, and 4) the condition that the corner is not observed at the beginning of the shot, it is one shot that forms a soccer corner kick scene” For example, information that can specify whether each of 1) to 4) is satisfied can be used as the Layer 2 feature amount.

  The highlight extraction unit 306 compares the statistical model for Layer 1 held in the model holding unit 302 with the Layer 1 feature amount from the Layer 1 generation unit 303, and Layer 2 held in the model holding unit 302. Are compared with the layer 2 feature quantity from the layer 2 generation unit 305, and the comparison is comprehensively judged to extract highlights.

  FIG. 39 is a flowchart showing an example of processing of the highlight detection unit 15 having the functional configuration of FIG. 38, that is, a detailed example of highlight detection processing in step S5 of FIG.

  However, in the example of FIG. 39, the statistical models of Layer 1 and Layer 2 held in the model holding unit 302 are “highlights when the first condition is satisfied” and “second It is assumed that a model with the content “is a highlight when the condition is satisfied” is adopted.

  Specifically, for example, when a soccer corner kick is extracted as a highlight, as described above, 1) the number of non- “Long” shots before the “Long” shot is within a predetermined range, and 2 ) A condition that the total period (length) of non- “Long” shots is a length within a predetermined range can be adopted as the first condition.

  Also, for example, as described above, 1) a goal box is observed over several frames, 2) no center line is observed in front of the goal box, and 3) the goal box moves from both sides of the screen to the center. 4) The condition that no corner is observed at the beginning of the shot can be adopted as the second condition.

  In such a case, in step S71, the Layer1 feature value generation unit 303 generates a Layer1 feature value for the moving image content.

  In step S72, the Layer3 feature value generation unit 304 generates a Layer3 feature value for each frame in the shot for each shot in the moving image content.

  In step S73, the Layer2 feature value generation unit 305 generates a Layer2 feature value for each shot in the moving image content.

  In the example of FIG. 39, for the sake of convenience, the processes of steps S72 and S73 are executed after the process of step S71, but the order of the processes is not particularly limited. As is clear from the example of FIG. 38, the processing in step S71 by the Layer1 feature value generation unit 303, the processing in step S72 by the Layer3 feature value generation unit 304, and the processing in step S73 by the Layer2 feature value generation unit 305 are respectively This is because they are executed independently and in parallel.

  In step S74, the highlight extraction unit 306 determines a predetermined shot of the moving image content as a target shot.

  In step S75, the highlight extraction unit 306 determines whether or not the Layer1 feature amount satisfies the first condition for the shot of interest.

  If it is determined in step S75 that the layer 1 feature amount satisfies the first condition for the target shot, the highlight extraction unit 306 determines in step S78 that the target shot is at least part of the highlight. Thereafter, the process proceeds to step S79. However, the processing after step S79 will be described later.

  On the other hand, when it is determined in step S75 that the Layer1 feature amount does not satisfy the first condition for the target shot, the highlight extraction unit 306 determines that the Layer2 feature amount is the first for the target shot in step S76. It is determined whether or not the condition of 2 is satisfied.

  If it is determined in step S76 that the layer 2 feature amount satisfies the second condition for the target shot, the highlight extraction unit 306 determines in step S78 that the target shot is at least part of the highlight. Thereafter, the process proceeds to step S79. However, the processing after step S79 will be described later.

  On the other hand, if it is determined in step S76 that the layer 2 feature quantity does not satisfy the second condition for the target shot, the highlight extraction unit 306 determines in step S77 that the target shot is not a highlight. To do. Thereafter, the process proceeds to step S79.

  In summary, when the process of step S77 or S78 is completed, that is, when the determination as to whether or not the shot of interest is highlighted is completed, the process proceeds to step S79.

  In step S79, the highlight extraction unit 306 determines whether or not the target shot is the last shot.

  If it is determined in step S79 that the shot of interest is not the last shot, the process returns to step S74, and the subsequent processes are repeated.

  In other words, the loop processing of steps S74 to S79 described above is repeated as appropriate for some of the shots constituting the moving image content to determine whether or not it is a highlight. And if the recognition process to the last shot is performed, it will determine with YES by the process of step S79, and a process will progress to step S80.

  In step S80, the highlight extraction unit 306 extracts highlights based on the above-described recognition result.

  As described above, in the example of FIG. 39, when the Layer 1 feature amount satisfies the first condition, or when the Layer 2 feature amount satisfies the second condition, the highlight is recognized. That is, the so-called OR condition is employed in the example of FIG. 39 as the highlight qualification condition. However, the example of FIG. 39 is merely an example, and other recognition conditions, for example, a so-called AND condition, that is, when the Layer 1 feature amount satisfies the first condition and the Layer 2 feature amount satisfies the second condition. The condition that it is recognized as a highlight may be adopted.

  As described above, with reference to FIGS. 38 and 39, the highlight detection process for detecting a highlight from a new moving image content has been mainly described on the assumption that the statistical model is held in the model holding unit 302.

  Next, the statistical model generation / update process will be described with reference to FIGS. 40 to 44. FIG.

  The statistical model generation / update process is performed by the learning unit 301 in the highlight detection unit 15 of FIG. FIG. 40 shows a detailed configuration example of the learning unit 301.

  In the example of FIG. 40, the learning unit 301 is configured to include an attention area extracting unit 351 to a highlight learning unit 358.

  Each of the attention area extraction unit 351 to the shot classification unit 354 has basically the same configuration and function as each of the attention area extraction unit 11 to the shot classification unit 14 of FIG. Each of the Layer 1 feature value generation unit 355 to the Layer 2 feature value generation unit 357 is basically the same as each of the Layer 1 feature value generation unit 303 to the Layer 2 feature value generation unit 305 of FIG. 38 of the highlight detection unit 15 of FIG. In general, they have the same configuration and function.

  That is, various videos of a scene to be highlighted, for example, a soccer corner kick scene, are input to the learning unit 301 as learning videos, respectively, to the attention area extraction unit 351 of the learning unit 301. Then, each of the attention area extraction unit 351 to the Layer2 feature value generation unit 357 executes the series of processes described above. As a result, the Layer 1 feature value and the Layer 2 feature value for the learning video are input to the highlight learning unit 358.

  Accordingly, the highlight learning unit 358 generates / updates a statistical model for Layer 1 by learning each Layer 1 feature amount for various learning videos, and causes the model holding unit 302 to hold the statistical model.

  Specifically, for example, the highlight learning unit 358 analyzes a shot type association rule between shots to obtain a state transition probability model, for example, an HMM (Hidden Markov Model) as shown in FIG. It can be held in the model holding unit 302 as a statistical model for Layer1. In FIG. 41, S1k (K is any integer value from 1 to 4) indicates each state of the HMM, and each of them is a predetermined one shot type, that is, in the present embodiment, “ Corresponds to "Long", "Medium", "Close-up", or "Out". That is, the probability of the transition of each state (shot type transition) and the probability of the occurrence of the state (corresponding shot type) can be obtained by learning each Layer 1 feature amount for the various learning videos described above. Become.

  Similarly, the highlight learning unit 358 generates / updates a statistical model for Layer 2 by learning each Layer 2 feature amount for various learning videos, and causes the model holding unit 302 to hold the statistical model.

  In this case, an HMM can also be adopted as a statistical model for Layer 2, and a frame type can be adopted as one corresponding to each state of the HMM. The type of frame to be used is not particularly limited, and a designer or the like can freely decide according to the scene to be highlighted.

  For example, when the above-mentioned soccer corner kick is extracted as a highlight, seven frame types as shown in FIG. 42, that is, “Middle”, “Goal Box”, “Upper Corner”, “Lower Corner” are displayed. "," Unknown "," Upper Goal ", and" Lower Goal "are preferably used.

  “Middle” refers to a frame type including a central white line 371. “Goal Box” refers to a frame type that includes the goal gate 372 and does not include the corner position 373. “Upper Corner” refers to a frame type that does not include the goal gate 372 and has a corner position 373 in the upper region. “Lower Corner” refers to a frame type that does not include the goal gate 372 and has a corner position 373 in the lower region. “Unknown” refers to a frame type that does not include the central white line 371, the goal gate 372, and the corner position 373. “Upper Goal” refers to a frame type including a goal gate 372 and having a corner position 373 in the upper region. “Lower Goal” refers to a frame type including a goal gate 372 and having a corner position 373 in the lower region.

  These seven frame types are determined for each frame by the Layer3 feature value generation unit 356 of FIG.

  Specifically, for example, the object feature extraction unit 352 attempts to extract the center white line 371, the goal gate 372, and the corner position 373 as object features from each of the frames constituting the learning video. The reason for “try” is that some frames may not be extracted. For example, the goal gate 372 is not extracted from the frame classified as “Upper Corner”.

  The Layer3 feature value generation unit 356 classifies each frame constituting the learning video into one of the seven frame types based on the extraction result of the object feature extraction unit 352, and the classification result. To the Layer 2 feature quantity generation unit 357 as a Layer 3 feature quantity.

  Note that the Layer 3 feature value generation unit 304 can be configured as a recognizer that detects the frame type of each frame, for example, classification learning such as SVM (Support Vector Machine), or a recognizer that uses a neural network or the like. In this case, the parameters of the recognizer can be obtained by statistical learning using each image sample.

  The Layer2 feature value generation unit 357 generates a time transition of the frame type of each frame in the shot that forms at least a part of the soccer corner kick scene as a Layer2 feature value, and provides it to the highlight learning unit 358.

  The highlight learning unit 358 can obtain an HMM as shown in FIG. 43, for example, and can store it in the model holding unit 302 as a statistical model for Layer2. In FIG. 43, S2m (m is any integer value from 1 to 8) indicates each state of the HMM, and each of the predetermined frame types, that is, this embodiment is shown in FIG. It corresponds to any of “Middle”, “Goal Box”, “Upper Corner”, “Lower Corner”, “Unknown”, “Upper Goal”, “Lower Goal”. That is, the probability of the transition of each state (frame type transition) and the probability of the occurrence of the state (corresponding frame type) can be obtained by learning each Layer 2 feature amount for the various learning videos described above. Become.

  Specifically, for example, when a state transition diagram as shown in FIG. 44 is obtained as a result of learning by the highlight learning unit 358, the state holding diagram is used as a statistical model for Layer 2 as a model holding unit. 302 can be held.

  The detailed example of the highlight detection unit 15 in the image processing apparatus of FIG. 1 will be described with reference to FIGS. By adopting such a highlight detection unit 15, it is possible to realize a highlight detection method designed based on rule analysis of the relationship between shot types and the relationship between frames. It is possible to easily extract various scenes as highlights.

  By the way, the series of processes described above can be executed by hardware or can be executed by software. When a series of processing is executed by software, a program constituting the software executes various functions by installing a computer incorporated in dedicated hardware or various programs. For example, it is installed from a program recording medium in a general-purpose personal computer or the like.

  FIG. 45 is a block diagram showing an example of the configuration of a personal computer that executes the above-described series of processing by a program. That is, all or a part of the image processing apparatus of FIG. 1, for example, some functional blocks can be configured as shown in FIG.

  In FIG. 45, a CPU (Central Processing Unit) 401 executes various processes in accordance with a ROM (Read Only Memory) 402 or a program stored in a storage unit 408. A RAM (Random Access Memory) 403 appropriately stores programs executed by the CPU 401 and data. These CPU 401, ROM 402, and RAM 403 are connected to each other via a bus 404.

  An input / output interface 405 is also connected to the CPU 401 via the bus 404. Connected to the input / output interface 405 are an input unit 406 made up of a keyboard, mouse, microphone, etc., and an output unit 407 made up of a display, a speaker, and the like. The CPU 401 executes various processes in response to commands input from the input unit 406. Then, the CPU 401 outputs the processing result to the output unit 407.

  The storage unit 408 connected to the input / output interface 405 includes, for example, a hard disk, and stores programs executed by the CPU 401 and various data. A communication unit 409 communicates with an external device via a network such as the Internet or a local area network.

  A program may be acquired via the communication unit 409 and stored in the storage unit 408.

  The drive 410 connected to the input / output interface 405 drives a removable medium 411 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, and drives the program or data recorded therein. Get etc. The acquired program and data are transferred to and stored in the storage unit 408 as necessary.

  As shown in FIG. 45, a program recording medium that stores a program that is installed in a computer and is ready to be executed by the computer includes a magnetic disk (including a flexible disk), an optical disk (CD-ROM (Compact Disc-Read). Only Memory), DVD (Digital Versatile Disc), a removable medium 411 which is a package medium composed of a magneto-optical disk, a semiconductor memory, or the like, or a ROM 402 in which a program is temporarily or permanently stored, or a storage unit 408 is constituted by a hard disk or the like constituting 408. The program is stored in the program recording medium using a wired or wireless communication medium such as a local area network, the Internet, or digital satellite broadcasting via a communication unit 409 that is an interface such as a router or a modem as necessary. Done.

  In the present specification, the step of describing the program stored in the program recording medium is not limited to the processing performed in time series in the described order, but is not necessarily performed in time series. Or the process performed separately is also included.

  Further, in this specification, the system represents an entire apparatus or circuit constituted by a plurality of apparatuses or circuits.

It is a functional block diagram which shows the functional structural example of the image processing apparatus to which this invention is applied. It is a specific example of the image processing which the image processing apparatus of FIG. 1 performs. It is a figure which shows the example of the process result of step S1 of FIG. It is a figure which shows the example of the process result of step 2 of FIG. It is a figure which shows the example of the process result of step 2 of FIG. It is a figure which shows the example of the process result of step 2 of FIG. It is a figure which shows the example of the process result of step 2 of FIG. It is a figure which shows the example of the process result of step 3 of FIG. It is a figure which shows the example of the process result of step 4 of FIG. It is a figure explaining the example of the processing content of step 5 of FIG. It is a figure explaining the example of the processing content of step 5 of FIG. It is a functional block diagram which shows the detailed example of a functional structure of the attention area extraction part of FIG. It is a figure explaining the HSV histogram model hold | maintained at the HSV histogram model holding | maintenance part of FIG. It is a figure explaining the HSV histogram model hold | maintained at the HSV histogram model holding | maintenance part of FIG. It is a figure which shows the specific example of the process result of the attention area extraction part of FIG. It is a figure which shows the specific example of the process result of the attention area extraction part of FIG. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining an example of the object feature extracted by the object feature extraction part of FIG. 1, and its extraction method. It is a figure explaining one method of shot cut detection by the shot cut detection part of FIG. 3 is a flowchart for explaining a detailed example of processing executed by the shot cut detection unit of FIG. 1, that is, the shot cut detection processing of step S3 of FIG. It is a figure explaining the kind of shot classified by the shot classification part of FIG. 1, ie, an example of a shot type. FIG. 31 is a flowchart for explaining a detailed example of shot classification processing in step S4 of FIG. 2 when the shot type of the example of FIG. 30 is adopted. It is a figure explaining an example of the information utilized by the shot classification process of FIG. It is a figure explaining an example of the information utilized by the shot classification process of FIG. It is a figure explaining an example of the information utilized by the shot classification process of FIG. It is a figure explaining an example of the information utilized by the shot classification process of FIG. It is a figure explaining an example of the information utilized by the shot classification process of FIG. It is a figure explaining the basic concept of the highlight extraction by the highlight detection part of FIG. It is a functional block diagram which shows the detailed example of a functional structure of the highlight detection part of FIG. 39 is a flowchart for describing a detailed example of processing executed by the highlight detection unit of FIG. 38, that is, highlight detection processing in step S <b> 5 of FIG. 2. It is a functional block diagram which shows the detailed example of a functional structure of the learning part of FIG. It is a figure explaining the model obtained as a result of the learning by the learning part of FIG. 40, and the learning. It is a figure explaining the model obtained as a result of the learning by the learning part of FIG. 40, and the learning. It is a figure explaining the model obtained as a result of the learning by the learning part of FIG. 40, and the learning. It is a figure explaining the model obtained as a result of the learning by the learning part of FIG. 40, and the learning. It is a block diagram which shows the structural example of the personal computer as an image processing apparatus with which this invention is applied.

Explanation of symbols

  11 attention area extraction section, 12 object feature extraction section, 13 shot cut detection section, 14 shot classification section, 15 highlight detection section, 51 HSV histogram model holding section, 52 down sampling section, 53 HSV component peak detection section, 54 attention Region range determination unit, 55 attention region detection unit, 56 post processing unit, 301 learning unit, 302 model holding unit, 303 Layer1 feature generation unit, 304 Layer3 feature generation unit, 305 Layer2 feature generation unit, 306 highlight extraction , 351 attention area extraction unit, 352 object feature extraction unit, 353 shot cut detection unit, 354 shot classification unit, 355 Layer1 feature generation unit, 356 Layer3 feature generation unit, 357 Layer2 feature generation unit, 358 highlight learning Department, 40 1 CPU, 402 ROM, 408 storage unit, 411 removable media

Claims (9)

In an image processing apparatus that performs image processing on an image obtained as a result of photographing a real-world subject,
Analysis according to a predetermined analysis method with a predetermined one of one or more real world backgrounds for a real-world subject as a target background and a plurality of images including a target region corresponding to the target background. Is held in advance, and holding means for holding image characteristic information indicating the image characteristic of the region of interest obtained as a result of the analysis, and
Analyzing means for performing analysis according to the predetermined analysis method on the processing target image, using the newly input image as the processing target image;
An image processing apparatus comprising: a detection unit configured to detect the region of interest included in the processing target image based on an analysis result by the analysis unit and the image characteristic information held in the holding unit. The predetermined analysis method is a method of analyzing a 3D HSV (3-Dimensions Hue Saturation Value) histogram of an image,
The image processing apparatus according to claim 1, wherein the image characteristic information is information indicating characteristics of the 3D HSV histogram of an image including the region of interest. The image characteristic information is information indicating each dispersion range of H, S, V that can be taken by the image including the region of interest,
The analysis means calculates the 3D HSV histogram for the processing target image, and detects the peak of each component of H, S, V from the 3D HSV histogram,
The respective peaks detected by the analyzing means are set as the respective center values of H, S, and V, and the respective dispersion widths of the H, S, and V held in the holding means are provided. , V, the image processing apparatus further comprises a determining means for determining the range of interest as a region of interest range,
The detection means detects, as the attention area, an area composed of pixels having H, S, and V component values within the attention area range determined by the determination means in the processing target image. 2. The image processing apparatus according to 2. The image processing apparatus according to claim 1, further comprising: an object feature extraction unit configured to extract an object feature from the inside of the attention area detected by the detection unit or an area in contact with the target area in the processing target image. The image processing apparatus according to claim 4, wherein the object feature extraction unit detects a boundary of the attention area from the processing target image, and extracts information indicating the detection result as one of the object features. The processing target image is an image obtained as a result of shooting a soccer game,
The region of interest is a region corresponding to a real-world soccer playfield,
The object feature extracting means further includes two line segments in contact with the boundary of the region of interest and one line connecting the two line segments as an area corresponding to a goal in the real world of soccer from the processing target image. The image processing apparatus according to claim 5, wherein an area that is surrounded by a line segment and that satisfies a stipulated dimension of the goal is detected, and information indicating the detection result is extracted as one of the object features. . The object feature extraction means further detects a region corresponding to a corner position in the real world of soccer based on the boundary of the attention region, and extracts information indicating the detection result as one of the object features. Item 7. The image processing apparatus according to Item 6. An image processing method of an image processing apparatus for performing image processing on an image obtained as a result of photographing a real world subject,
Analysis according to a predetermined analysis method with a predetermined one of one or more real world backgrounds for a real-world subject as a target background and a plurality of images including a target region corresponding to the target background. Is performed in advance, image characteristic information indicating the image characteristic of the region of interest obtained as a result of the analysis is held in the image processing device,
Using the newly input image as a processing target image, performing analysis on the processing target image according to the predetermined analysis method,
An image processing method including a step of detecting the attention area included in the processing target image based on the analysis result and the held image characteristic information. A program that causes a computer to control image processing performed on an image obtained as a result of photographing a real-world subject,
Analysis according to a predetermined analysis method with a predetermined one of one or more real world backgrounds for a real-world subject as a target background and a plurality of images including a target region corresponding to the target background. Is performed in advance, provided that the image characteristic information indicating the image characteristic of the region of interest obtained as a result of the analysis is given to the computer,
The computer is
Using the newly input image as the processing target image, performing analysis according to the predetermined analysis method on the processing target image,
A program including a step of detecting the region of interest included in the processing target image based on the analysis result and the given image characteristic information.

Images