| With the development of network technology,society is stepping into the electronic age.Videos have become an important tool for people to save and transmit information.People can get and share all kinds of information through videos.However,some sensitive information is spread through video,which will do great harm to the security of society.Compared with the sensitive information transmission mode that only takes pictures or characters as the carrier,it is more difficult to detect and intercept the information transmission of videos carrier.The goal of this thesis is to detect and cover the sensitive text information in videos to prevent the transmission of related bad information.Compare with that existing literature which study character recognition or sensitive information detection alone.In this thesis,text recognition and sensitive information detection are skillfully combined,and the actual processing of sensitive information in videos is based on character recognition.In this thesis,network videos are taken as the research object including several short videos of about 1 minute and long video of several hours,and the samples of recognition-classification-coding are carried out.The main research contents include:In this thesis,PP-OCR is used for text recognition of video images.When the shooting angle of images is too inclined,the distance is too far,and the light is too strong or too dark,the character recognition effect of this method will be affected.In order to reduce the influence of these interference factors,pretreatment processes such as gray processing,denoising and cutting are added to sample images before PP-OCR recognition in this thesis.These process can enhance the text features of images,reduce the recognition error,then obtain recognized texts of PP-OCR more completely.Based on the results of text recognition,Naive Bayesian is used to classify texts,and then sensitive videos are extracted.Because the traditional word segmentation method will cut all types of words in the text in the process of classification,which results in long time-consuming and heavy workload.In order to solve these problems,sensitive database is used for text segmentation,which only cut sensitive words in this thesis.It can reduce the number of terms in the classifier test set and shorten the running time.For sensitive videos,Open CV+local positioning keywords are used for information encryption,so as to achieve the result of cover of sensitive information in sensitive videos.The experimental results show that the recognition accuracy of the improved PP-OCR model for CTW dataset is increased 28%,from 60% to 88%.After using sensitive database to segment a large number of sample data,the sample size of the classification model is reduced and the running time is obviously shortened.The running time is obviously shortened,and the running speed increased by 11.52% on average.This model is effective in detecting sensitive images,and the classification accuracy and recall of the model has reached above 97%.Finally,the information coverage also achieved good results. |
