Recognition of Cheating Behaviors Based on Finetuning of Model Parameters

There are many problems with the current recognition methods of test cheating behaviors, namely, low accuracy, poor efficiency, and imbalance between positive and negative samples. To solve the problems, this paper proposes a classification and recognition method for test cheating behaviors through the transfer learning of pretrained models. Firstly, cheating samples, which mainly cover three cheating behaviors (peeking, passing notes, and checking cellphone) were collected from surveillance videos of exam rooms. The samples were enhanced through size transform and image synthesis. Next, multiple strategies were adopted to freeze the feature weights of the convolutional layers in the Darknet, before retraining the cheating classifier. In this way, a classification and recognition model was obtained for cheating behaviors. The model was tested on a self-designed dataset of test cheating behaviors. The results show that our method recognized 95.57% of cheating behaviors accurately, which is much better than the accuracy of the other methods. The realtime performance and accuracy of our method meet the application requirements.