Research And Application Of Examination Abnormal Behavior Recognition Based On Deep Learning

Examination is the main method of selecting talents.In order to ensure the fairness of the exam,invigilation is usually the way of sending special personnel and combining video surveillance,but reviewing a large number of surveillance videos manually is inefficient and will miss a large number of targets.Therefore,it is the goal of this paper to let the computer accurately and automatically identify the abnormal behavior of the exam.It is a typical target detection task to identify the abnormal behavior category and location of candidates in the exam monitoring screen.The traditional target detection algorithm has the problems of difficult to design features and poor algorithm performance,and deep learning target detection algorithms effectively avoid artificial design features extraction methods and diversified network structures provide different solutions for detection.This paper mainly does the following two aspects of research work:(1)Based on the deep learning target detection algorithms Mobilnet-SSD and YOLOv3,the models are trained.Firstly the exam monitoring video is recorded from multiple angles,the video images are intercepted at 7-frame intervals and the abnormal behavior of the candidates is marked,and the data set file in tf Record format is obtained after format conversion.Then this article analyzes the SSD algorithm process and optimization strategy and replaces the SSD backbone network from VGG-16 to the structure of the lightweight network Mobile Net,to avoid the problem that the model is too large to run on embedded devices.By comparing the characteristics of YOLO series algorithms,the YOLOv3 algorithm with more balanced detection speed and accuracy and better detection of small objects is chosen to train the model;finally,the transformation trend of the loss function during the training of the two models is showed,and m AP is used as the performance indicator to complete the evaluation and analysis of the model.(2)In order to improve the real-time detection of the exam abnormal behavior recognition model on ordinary performance machines,this article first uses Tensorflow and Python-OpenCV to write a detection program that can run successfully,and then proposes two real-time improvement schemes: frame alternation detection and dual-thread detection.Comparing the results of the three detection experiments of frame alternating single thread,frame-by-frame dual thread and frame alternating dual thread,it proves that the frame alternating dual thread detection strategy can maximize the real-time detection while ensuring that the program runs with low memory and consistent screen playback.The innovation of this article is:(1)Using deep learning target detection algorithm to construct an exam abnormal behavior recognition model with high detection accuracy through comparative experiments.(2)The use of frame alternating dual-thread detection strategy improves the real-time detection of the model on ordinary performance machines.