Research On Webcast Pornography Detection Based On Deep Learning

As webcast industry grows,the security problem of it is becoming increasingly obvious.At present,supervision of each broadcast platform almost relies on manual work of supervising staff and user reports.Nevertheless,large quantity of online broadcast room run at a consistent time,thus the human supervision is overwhelmed.Given above facts,this paper will present a porn detection model,which is based on deep learning.Compared with traditional neural network,deep learning could effectively solve problems of gradient diffusion and over-fitting,and has a better performance than traditional machine learning methods in the fields of image recognition,speech recognition and text analysis.Not only the image of the host but also the barrage sent by audience exists in the process of live streaming.Thus,in this paper,porn detection will be divided into two sectors.The first part tests the contents of live streaming based on deep learning.Then,the second part focuses on the barrage real-time transmitted by the audience in room.Video porn detection model M-ACORDE(Motion-ACORDE)is an optimization of the ACORDE(Adult Content Recognition with deep Neural Networks)model.ACORDE model could extract the features of key frames via convolutional neural network.However,key frames only contain the space information of the video.Therefore,ACORDE model will skip the action information of the video and misjudge indistinguishable videos such as breastfeeding,fighting and beach video.To the contrary,action features of videos could be detected by M-ACORDE model.M-ACORDE model is able to extract active information via dense optical flow algorithm and extract space feature and action feature from key frame and optical flow image via CNN.Finally,two features are combined and LSTM will be applied to classify videos.Experiments prove that M-ACORDE is better than ACORDE in sensitive video classification and enjoys a lower false positive rate in classifying indistinguishable videos.Porn barrage detection model could divide barrages into porn and no-porn type,and then classify barrages through LSTM.Input barrages will be divided into two types first.The forbidden words will be removed from the porn part and this part will be classified through the trained LSTM model finally.It is expressed in experiments that the LSTM based porn barrage detection model,compared with LDA model,is more accurate and excellent in porn barrage detection.