Safety Helmet Detection Based on YOLOv7 With Super-Resolution Reconstruction

Abstract

In the construction industry, safety helmets have been extensively utilized to minimize head injuries caused by accidents in the workplace. However, some construction workers continue to work without wearing safety helmets. Conventional surveillance methods such as using video surveillance systems with manual inspections have limitations such as high costs and missed detections. Therefore, target detection methods based on deep learning can be applied to enhance the effectiveness and efficiency of the supervision of safety helmet usage. In this study, a combination of YOLOv7 and SR reconstruction using the ESRGAN architecture is employed to identify the use of safety helmets. The training process from the combination of ESRGAN and YOLOv7 models resulted in precision, recall, F1-Score, and mAP@0.5 of 0.8664, 0.8362, 0.851, and 0.8826 respectively. The proposed method achieves promising performance for automatically overseeing the use of safety helmets in construction areas.