Multimodal Video Scene Segmentation Algorithm Based On Support Vector Machine

Video scene segmentation is an important part of content-based video retrieval.It takes shots as the research object and divides similar shots into the same scene according to the relevance of shot content,so that a complete video can be divided into multiple logical story units.The current video scene segmentation methods use image features as the underlying features of the video,and do not fully consider the information contained in the video shot,which leads to the low accuracy of scene segmentation.Based on the full analysis of video content and in-depth study of video scene segmentation,this paper proposes a multimodal video scene segmentation algorithm based on support vector machine.The algorithm has high recall and precision,and can realize the fast and accurate segmentation of a variety of video sequences,Improve the accuracy of finding specific video clips in massive video and reduce the time cost of finding.The main contents of this paper are as follows:(1)Video data preprocessing.In this paper,aiming at the problem of using image features to represent the low-level features of video,which leads to the loss of shot content,we extract the low-level features of video through the idea of multimodal fusion.On the basis of the extracted image features,we extract the Mel frequency cepstrum coefficients to represent the audio features,and use the feature extraction method based on statistics,Word frequency and anti document frequency are selected to describe the text features of the video,and the extracted three types of data are fused by simfusion algorithm as the multimodal features of the video bottom layer.(2)Semantic concept detection.This paper constructs a semantic extraction model based on the classification idea of support vector machine in machine learning,selects Gaussian kernel function to solve the optimal classification surface between different types of data,and uses libsvm software package to construct several semantic classifiers to classify the semantic concepts corresponding to shot key frames,and counts the number of relevant shots returned by each semantic concept in the data set,The evaluation index is used to quantify the effect of semantic classification.(3)Video scene segmentation.In the video scene segmentation based on semantic concepts,this paper uses semantic overlapping shot chain algorithm to divide different scenes.According to the experimental results and analysis,due to full consideration of the inherent characteristics among multi-modes,compared with literature [35],this paper not only achieves better results in semantic concept detection,but also achieves better results in semantic concept detection.Moreover,higher recall and precision are obtained in scene segmentation of various video data.Experimental results of a variety of video data show that the recall rate and precision rate of the proposed algorithm reach 91.18% and 92.81%,2.36% and1.29% higher than that of the reference [35],and the comprehensive index reaches92.45%,1.84% higher than that of the reference [35].