Yolov5s-PSG: Improved Yolov5s-Based Helmet Recognition in Complex Scenes

Abstract

In the field of industrial safety, due to the existence of color, distance and other reasons in complex industrial environments caused by the helmet small target detection methods have the problem of misdetection and omission, and the Yolov5s model for real-time detection of helmets is not ideal. for the purpose of Improvement of extraction capabilities multi-scale features, improve the network’s focus on key features, and achieve an excellent balanced effect for accuracy and speed, the model Yolov5s-PSG based on Yolov5s improvement is proposed. First, the C3_PPA module improves the detection accuracy through a multi-branch feature extraction strategy and an attention mechanism. Second, the SimAM attention mechanism enhances the inspection accuracy by adaptively weighting the feature maps and thus improving the detection efficiency. Finally, the GSconv convolutional layer combines the SC module and the DSC module in a channel shuffling manner to infiltrate the information of the SC module within some of the information of the DSC module, which improves the detection speed without changing the accuracy.Yolov5s-PSG model outperforms the conventional target inspection in terms of accuracy, recall, map value, and loss function, where the accuracy is improved to 93.1%, recall to 91.2%, and map value to 94.6%. These findings are important for helmet inspection to find the optimal solution in terms of accuracy and speed for better helmet inspection in shop-floor factory environments.