Design And Implementation Of Video Content Recognition System Based On Multidimensional Features Extraction

Video content recognition is to get the theme of video through analyzing the video and get abstract expression of video.Video content service is the main business in information networks.An report from Cisco showed that the global Internet traffic will triple in the next five years,and that video services will achieve to account for 82%of users' Internet traffic.How to deal with such massive video content has beacome a urgent problem.For instance,recognzing the commercials in TV broadcasting,detection of illegal shots for video service platform.Video content recognition has a brief and result through analysis of video,and is more able to meet the development requirements of information networks with massive videos.The commercial contains abundant business information.In this paper,we designed a video content recognition system and used to identify commercial in TV broadcasting.Commercial is a very abstract description of video.In order to recognize the content of commercial,it is necessary to detect shot boundary,classify commercial shot and non-commercial shot,then recognize commercial,because there may also be a trademark or slogan in a non-commercial shot.In the existing work,the efficiency of the shot boundary detection algorithm is poor,and the method of commercial shot classification can not meet the variety of the video type.We designed and implemented a video content recognition system which integrates multi-dimensional information and applies deep convolution neural network and machine learning method,meanwhile uses traditional computer vision processing technology.This paper has three specific contributions as follows.(1)A new method of shot boundary detection is proposed based on our experimental data.We acquired a whole day video of TV broadcasting for experiment,and designed a new method of shot boundary detection with color feature of frames.Compared with existing works,our method is simple to calculate,which can reduce the interference of shots gradient transition and facilitate subsequent shots classification and content recognition.(2)A new method of shot feature extraction is proposed based on convolution neural network,and uses these features to classify shots into two categories:commercials and non-commercials.Traditional video features such as color,texture,etc.can only describe the detail and low dimensional feature.We used convolution neural network to extract the features of the key frames and merge them into the shot feature.Compared with traditional features,this feature has higher dimensional abstract expression ability,and the network model used has high computing efficiency.Using this feature,the support vector machine classifier is trained to divide the shots into commercials and non-commercials,finally got precision 93.74%and recall 95.33%,respectively.(3)A video content recognition system is designed and implemented by integrating the features of image and audio.Most of the previous video recognition methods only use the image features,or fuse image and audio in feature level.We used the image features and audio features for video content identification respectively,and fused them in the results level to enhance recognition accuracy of shot.Based on that,the recognition results are promoted to commercial level,finally achieved precision 90%and recall 98.02%of video recognition.The method proposed in this paper has a strong generalization ability.It can be applied to video classification tasks such as illegal shots detection,if corresponding training data were applied.