Research And Lmplementation Of The Detection Of Network Image Sensitive Text

With the rapid development and application of the Internet,the threat and risk of network security become increasingly prominent.Some criminals use audio,words and pictures to carry out terror,obscenities,money laundering,reactionary,gambling and other activities on the Internet,which seriously affects the physical and mental health of the majority of Internet users.At present,the technology of sensitive information filtering in the form of pure text has been developed and widely used.For other transmission channels,relevant regulatory measures are still inadequate.In view of the transmission form of some sensitive words embedded in pictures,this thesis combines the new achievements of text detection and recognition of natural scenes,and proposes a sensitive text information detection method for network pictures.The main research work of this thesis includes the following three parts:(1)This thesis proposes a text positioning algorithm based on DG-Textboxes++.According to the different sizes of the network sensitive text images,this thesis optimizes feature extraction network and introduces the deformable convolution network,making the receptive field to respond to the word sequence changes.The redundancy of the detection frame is optimized and the comprehensive index F?mean of the algorithm in the ICDAR2015 data set reaches 69.3%.(2)This thesis proposes a text recognition algorithm based on DS-RCNN.Firstly,dense convolutional neural network is used to extract text features.Aiming at the irregular text of network pictures,the rectifying transformation network is used to horizontally correct the irregular text in pictures,and a database of synthetic text images is built to verify and test the recognition model of DS-RCNN.The word accuracy rate of the algorithm in the synthetic text images data set reaches 90%.(3)This thesis proposes a sensitive semantic detection algorithm based on BSMOTE.Aiming at the requirement accuracy of sensitive information detection on the Internet,this thesis designs a text semantic detection algorithm of sensitive information based on the first level keyword filter and the second level classifier.The sensitive words can be judged quickly by the sensitive word database.At the same time,in the case of unbalanced samples of sensitive information text,this thesis uses the second level text classifier based on BSMOTE algorithm to accurately detect sensitive information text.Relevant experimental results show that the sensitive semantic detection algorithm can effectively detect sensitive information text.Finally,a sensitive semantic detection system is built and tested by experiments,which proves the practicability of the system.