Research On Methods Of Video Semantic Feature Extraction

In the big data era based on multimedia information, social development has been impacted tremendously and human life also has changed significantly. However, it is difficult to deal with massive multimedia data issues such as analysis, searching and processing. Currently, retrieval as the main solution for these problems is to rely on manual processing which is time consuming. According to the above problems, researchers have studied object-based retrieval, motion-based retrieval and semantic clues based searching to evolve from the content retrieval to semantic retrieval, which aims at more intelligent, more accurate retrieval service.In order to solve above problems, it is necessary to remove redundant video frames in massive videos. The feature vectors are composed of several MPEG-7descriptors. And dissimilarity computation between frames can be gained by calculating the distance of the vectors. A new algorithm based on clustering, called RRBC (Remove Redundant Frame based on Clustering) is proposed to detect redundant video frames. The experimental results verify the effectiveness of the proposed algorithm in maintaining the timing sequence of video frames and the main content integrity of video. The proposed algorithm can also reduce data scale and computation for following steps.The features which include color, texture, shape, spatial relation and random sampling have been classified and analyzed based on the state of low-level video and image features extraction. According to the application requirement of infrared image, the algorithm FEIED (Filtered and Enhanced Image Edge Detection) is proposed which is based on image edge feature extraction by using Gaussian filter and Laplacian edge enhancement. The algorithm has been verified by lack detection of the experiment chalcogenide glass.After removing redundant video frames, data scale has been reduced and it is useful for extracting key frame of video. The methods of key frame extraction include four steps which are feature extraction, shot boundary detection, sub-shot boundary detection and key frame extraction. The method of MPEG-7descriptors was also used to calculate frame difference during feature extraction. Shot boundary detection used double threshold segmentation method which is typical one. To extract key frame on sub-shot can reflect the main content of video. Then a sliding window-based sub-shot segmentation algorithm SBDW(Sub shot Boundary Detection based on sliding Window)is designed. It is effective to keep the timing of sequence of video frames and it can transfer multiclass problems to several double-class problems which can reduce timing complexity and calculation amount. Moreover, a new key frame extraction algorithm KFES (Key Frame Extraction based on Sub shot) based SBDW is put forward. It transferred key frame extraction to the vertex cover problem.Video semantic feature extraction methods are studied based on research achievement. An innovative video semantic feature extraction method is proposed based on conditional random model, which include MPE (Method Parameter Estimation) and MR (Method Replace). It can achieve automatic extraction and label of video semantic feature for providing feasibility of semantic retrieval.According to the social application requirement of video retrieval, the video semantic feature extraction research could lead to several creative achievements. With consideration of fulfilling the increased retrieval requirements, more effective semantic retrieval framework needs to be constructed, which can provide more accurate semantic description of object, motion, scenes and event.