Multimedia grows rapidly due to advance devices like camera. So, Text query retrieval doesn’t fulfill user’s desire videos. So, this method can find the desire video by discovering temporal features, applying indexing and matching techniques. This method reduces the cost and increases the accuracy.
In this method, there is only three stage -
Preprocessing stage: In this stage, we divide the images into shots for shot detection. Then, we extract color layout and edge histogram from the selected shots for feature detection and these features help in shot clustering which is done by K-means algorithm. Each shot is represented by a symbol which belongs to its cluster number. Now, we apply the validation techniques for clustering quality. We set 3 thresholds for good clustering. Local proportion is number of shots in a cluster. Local density is a density of a cluster. Global density is a density of global clusters.
Indexing stage: In this stage, we build the FPI tree. First, we have to generate the temporal patterns. A subsequence has to slide along the shot sequence for considering the durations of shots is called window. Eg- There are 4 target clips. Each clips consist of several sequential shot-pattern. Clip 1 to 4 are {A,B,C,A}, {C,B,B,A,E,F}, {F,F,E,E,A,B,D,B,C,A,B} and {B,C,G,C,A,D,B}. Assume that a query clip has a set of pattern {B, A}. Every clip has this pattern but clip 4 will take large duration. The window size can be created static or dynamic. Two shot pattern(length of pattern) can match the sequence and reduce the size of tree. After generating 2 shot pattern within a sliding window, now FPI tree can be constructed. This is the FPI tree of that eg-
FPI tree is developed to speed up the video search without considering duplicates pattern.
Searching stage: Now we have to find out the most relevant video using FPI tree. We apply Depth First Search. If the target clip matches with queue by 2-pattern of a sliding window, then the count of the relevant clip is stored. At last, we get the counting table for target clips. So there may be some clips which have same count. For this, we have to re rank this result because sometimes patterns can be same but some features may be different.
Conclusion: This is a content based video retrieval by using pattern based indexing and matching techniques. This will give the high quality video efficiently.
References
- Adjeroh, D. A., Lee, M.C. A distance measure for video sequences similarity matching. In proceedings of IEEE conference on multimedia computing and system.
- Chen, L., & Chua, T. S. (2001). A match and tiling approach to content-based video retrieval. In Proceedings of IEEE international conference on multimedia and expo.
Comments
Post a Comment