Pornographic Information Detection Based On Key Points Of Human Body

With the rise of self-media,such as short videos and live webcasts,the Internet has a large amount of information and the content is uneven,and the supervision requirements for pornographic information cannot be met by human means alone,which has stimulated the development of methods for automatic detection of pornographic images.Traditional pornography detection algorithms use skin color detection or sensitive part classification of video keyframes and images to determine whether they are pornographic images.However,it is difficult to detect pornographic information that does not appear naked skin but contains sexually suggestive actions,such as twisting or touching on sensitive parts.In addition,the backgrounds of pornographic scenes in the film are changeable and complex,the human body interacts in a variety of ways and has different forms,resulting in many cases where the targets to be detected are blocked by each other or interfered with foreign objects.It is difficult to accurately extract all the motion characteristics of the human body.In order to solve the above problems,a novel pornographic information detection framework KP-ACD is proposed.Regarding the pornographic information that does not appear in the video with naked skin but contains sexually suggestive actions,KP-ACD judges pornographic content based on whether the movement features of the key point sequence of the human body appear in the sensitive area of the human body.The framework uses a multi-level sequential convolutional structure model to locate the key points of the human body,extracts texture information and spatial information and describes the constraint relationship between each part,and then predict the coordinates of key points of the human body that are interfered by foreign objects or obscured by the targets to be inspected,then remove redundant postures through posture elimination criteria,and obtain a set of unique human body key point coordinate sequences.On this basis,the affinity field is constructed to generate corresponding skeleton features,and after combining it with the image features extracted by Res Net50,the segmentation results of the human body instance are obtained through the segmentation module based on the region of interest and the sensitive areas of the human body are divided,providing information for pornographic information detection data support.KP-ACD uses LSTM to superimpose and fuse the human body’s key point motion trajectory with two additional spatiotemporal posture features.The sequence feature is the movement change in the sensitive area of the human body for pornographic content discrimination,so as to realize the pornographic information detection of video data.This paper evaluates and compares the performance of KP-ACD from three parts: human posture detection,human instance segmentation and pornography detection,in respect of the accuracy,the number of iterations,and effectiveness of each component.The experimental results show that the pornographic information detection framework based on the key points of the human body proposed in this paper can effectively detect pornographic information that does not show a large area of skin color but contains sexually suggestive actions in the video,and at the same time solves the problem of mutual occlusion or interference by foreign objects.Compared with the existing methods on Pornography-2k,it has a higher accuracy rate and basically meets the actual needs of pornographic information detection.