Improved Lightweight Helmet Wearing Detection Method For YOLOXs
Tao YaJie, Lian Pan · 2022
Aimed at the problems of slow check speed and low accuracy in the existing safety helmet wearing detection methods, an improved YOLOXs helemt wearing detection method was proposed. First, to be able to make the model’s overall construction lightweight and its size smaller, the YOLOXs backbone feature extraction network CSPdarknet is replaced with MobileNetV2, and the basic convolution procedure is the depthwise separable convolution. Secondly, in order to enrich the network feature information and make full use of the available network capacity, a lightweight ECA attention mechanism is added to the inverted residual structure of MobileNetV2. Finally, through knowledge distillation, the model is fine-tuned to obtain Mbv2E-YOLOXs. The experimental results on the public data set SHWD shows that compared to YOLOXs algorithm, the improved algorithm reduces the amount of model parameters and model calculation by 13% and 18.9% respectively; the average accuracy, safety helmet detection accuracy, and worker head detection accuracy increased by 7.6%, 9.25%, and 5.8% respectively, and the detection speed of the model reaches 60frame/s. Therefore, the improved YOLOXs model has good real-time performance and better accuracy for helmet wearing detection, which can meet the detection needs of practical application scenarios.