Research On Abnormal Behavior Detection Of Candidates Based On Deep Learning

In recent years,artificial intelligence has gradually entered people’s lives and is widely used in many fields.With the increasing call for the construction of standardized and efficient examination rooms,how to use artificial intelligence technology to improve the efficiency of examination rooms has attracted widespread attention.In the traditional way of invigilation,it not only needs to consume a lot of manpower and material resources,but also often fails to detect the cheating behavior of candidates,and the efficiency of invigilation is low.Therefore,in order to improve the efficiency of the examination room and maintain the fairness of the examination room,this paper proposes an abnormal behavior detection method based on deep learning for the possible abnormal behavior of the candidates in the examination room.In this paper,the abnormal behavior of candidates in the offline examination room is taken as the research object,and the YOLOv5 s model is selected as the basic detection model.In order to improve the detection accuracy of the detection model for the abnormal behavior of small targets,and at the same time make the model meet the needs of real-time detection,this paper proposes an improved scheme that integrates the CA(Coordinate Attention)attention module and the Mobile Net V3 lightweight network into the YOLOv5 s model,and retains the Focus structure idea to improve the detection accuracy and detection speed of the model.Since a large amount of sample data is required for model training,and no public data set of abnormal behavior of candidates has been found at present,this paper will organize personnel to record video of abnormal behavior of candidates,obtain video images by sampling and screening,and then use histogram equalization,image rotation and horizontal flip to obtain the original images.The data are expanded in three ways,and manually labeled by Label Img software to construct the sample data set required for training,which provides support for model training and testing.Compared with the experimental test,the detection accuracy of the improved YOLOv5 s model is increased by 1.4%,reaching 84.6%,and the number of inferred image frames per second is increased by 30.9%,reaching 78.8.The performance of the model has achieved the expected improvement and has certain application value.