Improved helmet-wearing detection method based on YOLOX

Wearing a helmet in industrial production is an important measure to protect the life safety of staff. Due to the complex construction environment and long shooting distance, the existing object detection methods have low detection precision and high missed and false detection rates for safety helmets. To solve the above problems, an improved YOLOX safety helmet wearing detection algorithm is proposed. A Dilated Spatial Pyramid is designed to acquire global information and expand receptive field. Three atrous convolutions with varying dilation rates are employed to connect the three dilated convolutions in series, thereby obtaining a larger receptive field and global information while minimizing the loss of feature information. Moreover, an adaptive attention mechanism is adopted to mitigate the loss of contextual information in feature maps and enhance the model's detection accuracy by accounting for the contextual information of feature maps. It can be seen from the experimental results of SHWD data set that the mAP value of the improved algorithm is 96.61%, and the detection precision of workers wearing helmets is 98.05%, and that of workers without helmets is 93.39%, and the detection speed can reach 38 frames per second. The high detection precision is realized under the condition that real-time detection can be achieved.