| Summary: | 碩士 === 義守大學 === 資訊工程學系碩士班 === 98 === Video completion is mainly used to repair the damaged region in video sequences. The main challenge is that the restored sequence should have the same visual quality as the original. In general, the conventional techniques calculate the pixel values around the damaged region and search the similar patches in whole video sequences to inpaint the damaged area. However, these approaches usually require very high computational cost, and therefore it is not suitable for the large region restoration.
In this paper, we propose an efficient system for video completion. The proposed method processes foreground object and background objects separately. For static background restoration, we will restore the damaged area in background model directly, and then apply the restored background to whole video. For foreground objects, we detect the moving object by subtract the background model, and then track the moving object using Kalman filter. For more robust completion, we also proposed an effective method to remove the object shadow. Furthermore, we build database with object and non-shadow object for matching and completion respectively. Finally, we presented an object matching technique to search the similar moving objects in the database and completion schedule for smooth completion.
The system works well for both foreground and background objects removal. Several experiments show that the proposed technique can be applied to special effects in video completion.
|
|---|
