Research And Implementation Of Violent Knife Holding Image And Violent Video Recognition System Based On Deep Learning

With the advancement of science and technology,massive amounts of information appear on the Internet.Since the Internet has brought convenience to people,some problems have also emerged,such as violent knife holding image and violent videos on the Internet,which will seriously affect the physical and mental health of the Internet users.in order to protect the physical and mental health of netizens,especially young netizens,and reduce the work pressure of the regulatory authorities,this thesis uses the current popular deep learning technology to analyze the violent knife holding image and violent video for identification in the network.The work done in this thesis includes the following aspects:Firstly,for recogniting the illegal image,this thesie first constructs a Violent Knife Holding Image Dataset(VKD).Due to the fact that violent knife images rarely appear in real life,images that meet the research have been collected through Internet crawling,and image enhancement technology has been used to expand the dataset,so that the constructed data set can provide data support for this study.Before processing these data,the transfer learning method is used for pre-training,and then the SENet is added to the Dense Net model to improve its performance.Through training,testing and comparation with other classic models,it will be compared with different algorithms of similar research.the model in this thesis has better effectiveness at accuracy,recall rate and F1 value and less running time by experiment.Second,for the violent video recognition,A lightweight spatio-temporal-channel attention(STCA)module based on R(2+1)D is proposed.This model includes two sub-modules,Channel-Temporal Attention(CTA)module and Spatial-Temporal Attention(STA)module,which make the network focus on important information along the spatial,temporal and channel dimensions.The robustness of the model is verified by several sets of experiments.Compared with the dual-stream network,3DCNN network,and R(2+1)D network on the RWF-2000 dataset,STCA-R(2+1)D model in the thesis shows better accuracy in several different dataset.Finally,a violent knife holding image and violent video recognition system is built to identify violent knife-holding images and violent videos on web pages.from the perspective of different users,the demand analysis of violent knife image and violent video recognition system is carried out.Based on this,the system architecture is designed comprehensively,moreover the system functions are realized.In terms of recognition functions,it includes the recognition function of images of violent knife holding and the violent video in web pages,There are some other system management functions,such as: web page data storage function,blacklist management function,regional heat map management function,user management function,role management function,etc.then,the system is tested and analyzed,including functional and non-functional aspects.From the results of the test,it is concluded that the system is reliable and meets the expected requirements.