A Real Time Mask Wearing Detection Based on Lightweight CenterNet in Complex Scenes

CenterNet is a one-stage target detector based on key points with high detection accuracy. However, its backbone is Hourglass network with a large number of parameters, the recognition speed is slow and cannot be recognized in real time. In this paper, we proposes a lightweight Hourglass network based on CenterNet model for mask wearing detection. Firstly, we adopts the depth wise separable convolution network in the reverse residual block of the Hourglass network. In the upsampling and downsampling block, different stride set and two branches are used to reduce the number of model parameters and improve the detection speed. Secondly, we redefine the focal loss function which can correlate the loss values of two Hourglass networks and complement each other to improve the accuracy of difficult targets in complex environments. Finally, in order to improve the test robustness of the method, we constructed a data set of masks under different challenge scenarios. The experimental results show that the average accuracy of our method is 0.922 and the parameters are reduced to 1/25 of CenterNet, and the detection speed is increased by nearly 3 times. Our method can achieve real-time mask wearing detection in videos with better robustness, which provides practicality for deploying the network model to mobile terminals.