＜Introduction＞
　For the contents, similarity estimation and sense based retrieval of images or videos, the category of the images or videos determine a appropriate retrieval method. In other words, there is no applicable retrieval technique for all category of images or videos. For example, it is necessary for detecting nature images containing a clump of trees or buildings to use their frequency feature, other cases, detecting artificial images(CG) to make good use of information of color scheme etc.. Therefore, application should change by the categoly of images or videos.
　At first in this paper, the video category is broadcasting of Japan Professional Baseball.
　For broadcasting video of Japan Professional Baseball, there is a measure of similar pattern of camera-works and video-switching by each kind of plays. Accordingly, it is possible to detect and classify each play-scene from video sourse.
　This paper proposes The Scene Rtrieval and Arrangement System that has several play-scene pass filters including each knowledge based on camera-works and composition of images.
　This system cannot only construct several play-scenes, but useful for auto scoring and viewers review play-scenes or score-data easily. Moreover, it will be able to skim by excepting unimportant scenes.

＜System Conseption＞
　Figure 1 shows the system conception, consists of the video server stored the broadcasted video source, the video analisis which is the core of this system, the sub-server to store the classified play-scenes and GUI to select usr's request scenes.

Figure1:System Conception

＜Play-Snene Retrieval＞
　Generally, one batting frame sequence's contents consists of, in order, Batter's Expression, Pitching, Bat Swing, Taracking a Batted(Throwing) Ball, and Runner's Appearance.
　In these scenes, an interval from Pitching Starting Point to Next Batter is needed to see plays or to conjecture strategy and tendency for viewers(users).
　Consequently, I design each play-scene pass filter.

* Cut Detection
　Once the system detects the cut point, using Inter-frame Differential (between two frames before and after). The cut point is defined when the differential value is over threshold.

* Pitching Cut Sequence Retrieval
　Each cut frame detected certainly is checked whether or not the frame is general conposition to grasp pitching. The judgment depends on the amount of differentials between the cut frame and sample pitching images in databese prepared beforehand. Incidentally, some samples are selected in obedience to studium information which user input in advance.
　All samples picked up are used in calculating differential, and then if each value of their differential are than less the threshold, the system detects as pitching cut sequence.
　Pitching Cut Sequence is defined as an interval from the point detected pitching cut to next cut point. Therefore, this interval concerns frames is unrelated throwing to catcher.

* REAL Pitching Scene Detection
　The system must detect REAL Pitching Scene from the Pitching Cut Sequence in order to skim video. REAL Pitching Scene means frame sequence containing the event that one pitcher throws to catcher.

• Appointing the region which in pitcher exists.
• In the pitcher's region, computing motion vectors of each macro block.
• An amount of level motion use in evaluation of pitching motion.(At Pitcher's backswing, a large volume of level motion is detected.)
• The interval through before 10 frames of the detected point to after 30 frames is defined as REAL Pitching Scene.

* Each Batting Scene Retrieval 　Other Play-Scene(Batting-Scene) pass Filters, by means of inferencing camera-work with optical flow, judge the place batted ball has gone and extract frame sequence casting the situation of tracking batted ball.

I shall have been made this research by next year......
You must not expect too much of me(my research). Maybe.....