| With the constant progress of video transmission and transcoding technology,network traffic has a trend of being dominated by the video.Varieties of violence,horror and other illegal special videos flooded the internet,seriously affect the physical and mental health of the netizen and become a public security hazard.Faced with large-scale video,how to detect special video content effectively has become an important research topic in the field of computer vision.Recent years,with the rise of deep learning,traditional feature descriptors such as HOG and SIFT are being gradually replaced,and deep features tend to become the mainstream feature characterization method.Deploying deep features such as audio,image,video,text and other content not only outperforms traditional methods in characterization,but also has tremendous advantages in processing ability as well as great room for improvement in performance.Therefore,it is of far-reaching significance that the effective use of deep features to solve practical problems is a topic that should be focused on now and in the future.This paper studies how to use deep features for detecting the violent content of video effectively.The main work of this paper is based on three-dimensional convolutional neural networks(3D ConvNet).The main work includes the following aspects:(1)This paper studies the end-to-end detection algorithm based on 3D ConvNet,and analyzes the effect of the input frame length on the detection result.(2)This paper studies the pre-processing methods of 3D ConvNet,and proposes a modified frame sampling method.(3)This paper explores the algorithm framework of the combination of 3D ConvNet and different classifiers,and proposes a special video detection algorithm framework based on 3D ConvNet and extreme learning machines.Meanwhile,this paper compares test results of end-to-end 3D ConvNet,3D ConvNet with support vector Machine and 3D ConvNet combined with extreme learning machine. |
PDF Full Text Request