Lightweight Security Wear Detection Method Based on YOLOv5

Given a large number of network parameters of the existing security wear detection, it is difficult to run on the embedded platform. Based on the idea of deep separable convolution, a lightweight security wearable target detection network based on improved YOLOv5 is proposed. Specifically, the feature extraction network structure of YOLOv5 is lightweight improved to reduce computation of the proposed model, including the increase of the number of network layers and decrease of the number of parameters. In addition, an attention mechanism is introduced to weigh different channels of the feature map to improve detection accuracy. The model has been tested on PASCAL VOC dataset and security wear dataset. The experimental results show that the size of the proposed model is 8.0 MB, the number of parameters is 7.5∗105, and the number of FLOPs is 7.5∗105. Compared with the YOLOv5 model, the required memory is reduced by 44.8%, and the number of parameters decreased by 45.58%, FLOPs decreased by 54.54%. Accordingly, the results have demonstrated that the proposed method can significantly improve the detection speed while maintain the accuracy. Especially, we have successfully deployed the proposed model in the high-speed detection of security wear.