Research On Shot Retrieval Based On Fuzzy Evolutionary AiNet And Probabilistic Distance

With the rapid development of multimedia technology, multimedia data are increasing exponentially. Consequently, how to get the video data we need efficiently from abundant video databases becomes very important and urgent that Content-based Video Retrieval (CBVR) has become the research hotspot. In the process of Content-based Video Retrieval, the video data are divided into key-frames, shots and scenes by analyzing video structure. In the units of the shots, according to the user-submitted video examples, the similar video clips in the video database can be found and displayed in accordance with the degree of their similarity.In this work, we firstly discuss the background and then analyze the main existing expression content-based video retrieval algorithms. After analyzing the methods currently used by others, we propose a key-frame extraction algorithm based on fuzzy evolutionary aiNet and a shot similarity method measure based on probabilistic distance. The works are described as below:(1) key-frame extraction algorithm based on fuzzy evolutionary aiNet is presented. First of all, take every frames of the same shot as an antigen (Ag), and take the node in the fuzzy evolutionary aiNet (Artificial Immune Network) as an Antibody (Ab), which is the key-frame we want to extract. And then update the fuzzy evolutionary aiNet by clone compress and network compress. Finally, we can get the representative key-frame(s) of the shot in the fuzzy evolutionary aiNet.(2) In this paper, to represent the temporal features of a shot, a shot-weighted color histogram is defined, and each shot corresponds to a shot-weighted color histogram. In order to extract the spatial features of a shot, we construct a texture vector and a histogram of spatial structure information, where texture vector was constructed by computing the gray level co-occurrence matrix with five representative parameters of Entropy, contrast, energy, and inverse correlation gap, and histogram of spatial structure information was constructed by calculating area ratio, centroid, the standard deviation of x direction and y direction, the regional aspect ratio of x direction and y direction.(3) shot similarity method measure based on probabilistic distance is proposed. To calculate the spatial-similarity of two shots, at first, the spatial features were mapped to a high dimension space by nonlinear mapping, in which they obey Gaussian distribution. Then computed the probabilistic distance of spatial features in high dimension space, the spatial-similarity was computed by weighting spatial feature probabilistic distance. While the temporal similarity of two shots was obtained by histogram intersection of the corresponding shot-weighted color histograms. At last, the shot similarity was calculated by weighting temporal similarity and spatial similarity.(4) Construct a framework of video retrieval prototype system. A prototype system of content-based video retrieval designed and implemented by using the Modularization-oriented methods. It can be used to prove the effectiveness of the above proposed algorithms.