Research On Detection Technology Of Pornographic Information Based On Deep Learning

With the rapid development of the Internet,the content of the Internet has been diver-sified,complicated,and quantified,and all kinds of bad information(especially obscene pornography)are flooding in the Internet.In order to protect the physical and mental health of adolescents,different types of pornographic information on the Internet need to be fully supervised.Since the main forms of pornographic information include text,image,video,audio,therefore,this thesis carried out research on these types of porno-graphic information.The biggest problem of traditional detection methods is that single feature matching is easy to bypass,and deep learning algorithms have good feature ex-traction and feature fusion mechanisms.At present,deep learning has been widely used in computer vision,natural language processing,and speech recognition,and it has reached the most advanced level in these fields,so this thesis mainly use deep learning to identify pornographic text,pornographic images and pornographic videos.Pornographic text recognition usually uses keyword filtering,but this method is easy to bypass.Common bypass mechanisms include homophonic bypass,homomorphic by-pass,“pinyin”bypass,etc.These bypass mechanisms are widely available on the Internet.The cyber security is under great threat.In order to prevent these bypass mechanisms,this thesis uses four forms of Chinese characters,“Pinyin”,“Pinyin”initials,and“Wubi”coding as the input of the recurrent neural network model,using the attention mechanism,and then implement pornographic text recognition through feature fusion.After testing,the accuracy rate is as high as 99 %,and it can effectively prevent the above bypass mech-anisms,and it is more robust.Pornographic image detection sometimes has ambiguity,that is,the same picture has different definitions under different objects,events,and environments.Therefore,this thesis first discusses the definition of pornographic images in detail,and classifies porno-graphic images under given standards.There are four categories: absolute pornography,partial pornography,sexy,and normal.Aiming at this classification standard,a network model based on the combination of global classification and sensitive local classification is designed.Global classification mainly adopts a network architecture combining tradi-tional image features and convolutional features to achieve classification of global images.Sensitive local classification uses YOLO-like thinking to roughly determine whether there is sensitive information.Finally,the two classification information is stitched together to obtain the final classification result.After experimental testing,the accuracy rate is as high as 96.7 %,and the false negative rate is low.Pornographic videos can be viewed as a collection of porn images.This thesis mainly solve the problem of high complexity of porn video processing,adopts the clustering method to obtain key frames,and uses depthwise separable convolutions instead of ordi-nary convolution to reduce the amount of calculation.In order to unify into an end-to-end training method,a three-dimensional convolutional network is used to obtain spatial and motion features,an MFCC algorithm is used to extract audio information,and a BLSTM is used to obtain audio features.Finally,all features are merged in a bitwise manner to obtain the final classification result.After experimental tests,the accuracy rate is as high as 95.5 %,and the method has better detection effect when the audio signal is strong.