Real Time Recognition Of Sensitive Words In Video Stream

With the rapid development of mobile Internet,the network provides a large number of knowledge learning materials for the majority of young people,but also brings a threat to the healthy growth of young people that can not be ignored.It is very easy to spread highly sensitive information such as pornography,violence and reactionary language on the Internet.Many outlaws spread bad information by embedding words in videos.Compared with the simple sensitive information in the form of text type,it is hidden and difficult to be detected in this way.The data source of sensitive video can be not only the video of Internet website without audit,but also the video uploaded on the enterprise internal network.The goal of this paper is to design a system to solve the problem of sensitive text detection in website video.This paper improves the yolov3 network according to the text features,obtains the prior box size through the K-means clustering algorithm,and adjusts the network structure and pooling module to realize the text positioning function;The structure of encoding and decoding based on attention mechanism is adopted to realize the function of character recognition;Chinese word segmentation tool is used to segment the text,and two sensitive detection functions are used to filter the text after segmentation.The first filtering function preliminarily filters the text through prefix tree,and the second filtering function uses traditional SVM to detect the semantic of the text represented as sensitive statement by the first filtering function.The test results show that the proposed method not only meets the requirements of accuracy and real-time,but also can complete the detection of sensitive words in video stream.