Research On Video Emotion Recognition Based On Multi Feature Fusion

With the rapid development of digital media processing technology and computer network technology,as well as the continuous popularization of intelligent equipment,digital images,audio and video have gradually become main ways of information dissemination.Video has been widely applied because it is intuitive and vivid,and there are a large number of videos shared on the Internet every day.In order to organize and manage these massive videos conveniently and effectively,we need a technology of classification and retrieval based on the semantic content of video.Video content includes not only events and movements,but also abundant emotional semantic information,thus it is significant in research of the video emotional semantic,it improves the efficiency of retrieval and annotation of digital media,as well as the development of emotional interaction ability of digital entertainment products and artificial intelligence.Emotion is one of the inherent characteristics of video.Video emotion recognition is mainly about analyzing the emotion contained in video or predicting which kind of emotional experience it will brings to the viewers.Taking user-generated video as research object,this article presents emotional identification of video from the perspective of emotional cognition which is based on methods of machine learning.The specific research work and main innovation points of this paper are as follows:(1)In the preprocessing of video emotion recognition,randomly grabbing video frames or using traditional clustering methods to extract key frames usually give rise to a large redundancy in frames and lack of representativeness,to solve this problem,a key frame extraction method based on distance threshold clustering is proposed in this paper.This method completes clustering by intensive sample sampling to determine the distance threshold,which not only improves the traditional threshold selection problem of clustering method,but also improves the validity of key frames and representational ability of video.(2)In the aspect of video emotion expression,there may be some important regions in the key frame,thus this paper proposes the Bag-of-Word model and the spatial Pyramid model to block and quantify the key frames to obtain more complete visual feature information.In addition to the use of audio-visual features to express the personalized emotional content of video,we introduce two deep semantic attribute features named Classemes and SentiBank,which are rich in semantics,it compensates the semantic gap between low level audio-visual features and high level human emotions,so that the emotional expression of video become more abundant.According to a large number of experimental results,the model based on attribute characteristics can produce excellent performance in emotion recognition.(3)Inspired by the influence of multi feature fusion on classification accuracy,a video emotion recognition method based on feature weight adjustment is proposed in this paper.Different classification models are constructed and feature weights are trained by different feature descriptors,and multiple classification models are combined to complete the ultimate classification.In addition,the article also discusses the multi feature fusion of multi kernel learning,and applies multi kernel learning to video emotion recognition.By processing the feature of linear fusion after the kernel function,the emotion recognition of the video is improved.