Research On Special Video Content Detection Algorithm Based On Gradient Direction Histogram Information

With the development of rich media content distribution technologies,a large number of videos are widely spread through mobile phone and internet.But videos' quality and content are diverse,many videos contain violent information,which will become a security hazard to society.Thus,violent video detection and filtering are of great significance and it will provide solutions to the scientific study of related field.This paper studied the detection of violent video,the main research contents are as follows:1.A new feature Conv-CoHOG3D based on Pooling operation is proposed.The problems of the high dimensional feature are difficulty in training,time-consuming,and the high requirement of the computer.The use of the Pooling operation can realize the dimensionality reduction of the high dimensional feature efficiently.And the pooled Conv-CoHOG3D feature still have high representational ability.2.A video violence content detection algorithm based on Conv-CoHOG3D and the bag of word model is proposed.The Conv-CoHOG3D feature is extracted from video clips and the visual dictionary is obtained by K-means clustering algorithm.Then the bag of word model about the video segment is constructed.And the word frequency vectors are processed in order to fit the requirements of the classifier.At last,we train model and test videos.3.Verifying the detection effect of different combinations of gradient direction histogram features and classifiers,finding the optimal classifier.The results can provide reference data and solution for related research.