Multi-feature-based Video Scene Classification

In recent years,the rapid development of network information technology has caused the rapid expansion of video information.This poses considerable challenges for fast video retrieval.In order to improve the efficiency of retrieving video,video scene classification technology can be used to accelerate video detection efficiency to achieve fast video retrieval.This paper studies the spatial location information of scene keyframes in scene classification and its segmentation,scene feature extraction and fusion,and scene classification.The main contents and innovations are as follows:(1)Video scene average key frame division.The video key frame contains rich video scene information,so the key frame set can be used to represent the main content of a video stream.For scene recognition,the spatial location information of the scene is very important for scene recognition.Considering the mutual spatial position relationships between scenes within the keyframe image and the different scenes' contribution to scene recognition,the video scene keyframes are divided into the interest regions(ROI)and sub-regions of interest are divided(SROI)and areas of interest(NOI).By dividing the scene average key frames,this largely eliminates content that is ineffective for scene recognition,and also ensures the integrity and relevance of the content of the identified scene.(2)Video multi-feature extraction and fusion.The scene recognition rate is related to its description features.In the scene recognition process,if only the unique features are used to describe the scene information,there will be a large deviation between the description quantity and the real quantity.Extracting various features of the scene can effectively solve the problem of scene description deviation.Color features and texture features are effective scene description features.In the experiment,the gray level co-occurrence matrix and "hue product"features of the average key frame are extracted,and these two features are weighted and parallelly merged to obtain the integrated features of the video scene.(3)A threshold decision classification algorithm was proposed.The feature threshold is used as a standard to measure the similarity between video scenes.For similar scenarios,the description of scene content has strong consistency.LDA classification and K-nearest neighbor classification were used as comparison experiments.The experimental results show that the threshold judgment algorithm has high accuracy and the accuracy rate reaches 80%,which is 11.25%higher than the LDA classification method and 5%higher than the KNN classification method.And the classification result of the threshold judgment classification method has a strong stability and is a robust classifier.