Research On Campus Bullying Detection Algorithm Based On Video And Audio Joint Recognition

In this era of rapid development of network media technology,people have more and more channels to receive information,and the campus bullying incident has begun to enter people's vision.In the era of Internet self-media,the violent and vulgar information in the online world has a serious impact on the mental development of young people.Some students even imitate the violent and vulgar behaviors that they are exposed to.Campus bullying incidents not only affect campus life The atmosphere also has a bad influence on the psychological development of students,which shows the importance of actively detecting campus bullying incidents on campus.In this paper,the pattern recognition technology is used to judge the situation of students being subjected to campus bullying from both video and audio aspects,and the discrimination results of the two aspects are fused through improved DS data fusion.Based on the campus surveillance video combined with the directional audio acquisition equipment,the intermittent real-time monitoring of the physical and mental safety of students is realized,which is conducive to the construction of a harmonious campus environment.According to the complexity of video image data features,this article first preprocesses the video image data and builds a deep convolutional neural network to analyze bullying actions and daily actions to extract action features.The 16-frame video image data extracts a 4096-dimensional feature vector,and on this basis,a neural network recognition algorithm is designed.The recognition rules of the recognition algorithm are as follows: 16 frame video images are used as the basic recognition unit to segment the video sample data,and each basic unit is judged once,and finally the accuracy rate reached 95.65%,the recall rate reached 88.00%,and the F1-Score value was 91.67%,indicating that the classification model performed well.According to the complexity of audio data features,this paper firstly performs pre-emphasis,framing,and windowing preprocessing on the audio data,and secondly extracts the MFCC feature parameters from the audio data.Based on the successful extraction of feature parameters,a deep convolutional neural network was built to design a violent emotion recognition algorithm.The design of the violent emotion recognition algorithm in this paper is based on the self-made small voice database,the Finnish voice database and the CASIA public voice database.For the self-made small voice database,the accuracy rate of the recognition algorithm is 88.33%,F1-Score is81.14%;for the Finnish voice database,the accuracy rate of the recognition algorithm is95.00%,F1-Score is 95.00%;for CASIA open voice database The accuracy rate of therecognition algorithm is 91.67%,and the F1-Score is 91.43%.The violent emotion recognition algorithm in this paper shows good performance on the three speech databases,which proves the versatility of the algorithm.Finally,in view of the limitations of the DS fusion algorithm,a new fusion rule is proposed.And the improved DS fusion algorithm is used to fuse the recognition results of both video and audio.The accuracy rate is 94.33%,and the F1-Score is 94.07%.Compared with the improved DS fusion algorithm,the accuracy of the algorithm is improved.10.19%,F1-Score increased by 2.66%.