Abstract: In order to improve the detection accuracy of convolutional neural networks in student behavior recognition applications, this paper uses K-means clustering to cluster the unique data sets to obtain more adaptive anchor box, and proposes a YOLOv3 network based on improved loss function. The network model dynamically transforms the original squared loss function weights, focusing on the calculation of the loss of continuous variables. The new loss function can effectively reduce the influence of the gradient disappearance of the sigmoid function, making the model converge more quickly. The experimental results show that the deep convolutional neural network based on the improved loss function has improved the recognition of the three poses of “lookup”, “lookdown” and “talk”.

Key words:  , K-means； image enhancement； loss function； YOLOv3 network； gesture recognition