Safety helmet wearing detection in thermal power plant based on improved YOLOv5

In industrial safety accidents, the failure of workers to wear helmets is a major cause. A deep learning detection model can now be developed using computer vision techniques to detect whether workers in thermal power plants are wearing helmets. An improved helmet-wearing detection algorithm for YOLOv5 is proposed in this paper which overcomes the shortcomings of existing algorithms and successfully detects small, obscured, and dense targets. The algorithm enhances the efficiency of detecting small targets by adding classification convolutional blocks. Furthermore, it integrates the MA channel attention mechanism into the Neck network to reconstruct channel weight, filter out erroneous channels, increase feature expression capability, and introduce FReLU to enhance the spatial sensitivity of the activation function. Under the SHWD dataset, the mAP of this model reaches 91.9%, an improvement of about 1.6% as compared to the original model. The study demonstrates that the improved algorithm can satisfy the practical requirements of thermal power plant work sites and performs well for the helmet detection task.