Research On Key Technologies Of Video Classification And Retrieval

Video has become an important part of entertainment and social activities due to its vivid,intuitive and versatile property,but it has also led to a dramatic increase in video number.How to effectively classify and retrieve massive video data has become one of the research hotspots in the field of computer vision and information retrieval.Due to the complexity of video content and structure,the validity of video spatiotemporal feature representation has become a difficult point.This paper has carried out in-depth research on key technologies of video classification and video retrieval.The main work is as follows:1.A video representation algorithm based on multi-level pooling pseudo-3D convolutional neural network(CNN)is designed and implemented.Firstly,the spatiotemporal features is extracted by multi-level pooling pseudo-3D CNN.After that,the uniformly pooled features are sent to the fully connected module consisting of a fully connected layer and Softmax layer.And the maximum pooled features are passed through encoding module that consists of PCA,Vector of Aggregate Locally Descriptor(VLAD)and Support Vector Machine(SVM).2.A feature representation algorithm combining pseudo 3D CNN and 2D CNN is proposed.The shortcomings of image frames in video representation are compensated by motion vectors.The experimental results on several data sets verify the effectiveness of the algorithm.3.An effective multi-stream information fusion algorithm is proposed to improve the robustness of the video classification algorithm.4.This paper adopts the combination of pseudo 3D CNN,feature dimension reduction and KD-Tree retrieval,which effectively reduces the complexity of the retrieval algorithm and improves the efficiency of the retrieval algorithm.Moreover,the experimental results also show that its performance is better than the mainstream hash-based search structure.5.A Web-based demonstration system is built for the video retrieval algorithm of this paper.