Detection Based On Wavelet Neural Network Lens

In recent years, the content-based video retrieval (CBVR) is becoming one of the most active research topics in the field of information indexing in the world. Shot boundary detection (SBD) is the key technique of CBVR. Unfortunately, there is no complete and reliable system for the complexity of itself. This dissertation focuses on the methods of SBD based on the wavelet transform and neural network, and some useful works have been done on the lax-model wavelet network of SBD. The main research fruits achieved in this thesis are given as follows:An improved histogram-difference-based method with tolerance for SBD is realized. First, the two-dimensional wavelet transform representations of these frame images are obtained. Second, the low-frequency parts of them are changed from the RGB space into HSV space. Third, the differences of neighbor frames are extracted through the blocked histogram. At last, integrate the window and dual-threshold methods to determine shot boundary. Adjust the adaptive threshold parameters and add the detection tolerance into the method based on the experimental analysis of SBD.One kind of non-threshold value SBD method, which based on neural network, is used to eliminate the difficult of choose thresholds or parameters for the detection of different type videos. On the foundation of traditional histogram differences and pixel differences, characters of abrupt change are revealed effectively by computing twice frame difference. Then the integrating the non-neighboring frame difference and the neural network to detect the gradual changes, the experimental results shows this method is efficient.Because massive frames information is needed to carries on the gradual changes, and the speed is slow, the efficiency is not good. Therefore, the various sub-bands'information of the wavelet transform is used to detect shot changes respectively. First, the blocked histogram differences of low frequency information are obtained. Second the neighbor frame differences, the twice frame difference and the non-neighboring frames differences are used to detect abrupt change and gradual change by the neural network, then the edge of each high-frequency unit is calculated. The fades transitions and dissolve transitions are judged by neural network according to the total number of appeared and disappeared edge points. The cuts can be detected well, and the fade in, fade out, dissolve are also distinguished well.