Improved Mask R-CNN Method for Intelligent Monitoring of Helmet in Power Plant

Abstract

Traditional helmet wearing detection algorithms frequently encounter the problems of low accuracy and poor robustness during the processing of power plants monitoring image with complex background, low image quality, target occlusion, and small targets. A helmet wearing detection algorithm based on improved Mask R-CNN was proposed. By improving the backbone network structure of Mask R-CNN algorithm and increasing the number of anchors, the detection accuracy and robustness were improved correspondingly. Meanwhile, Multi-NMS algorithm was proposed to remove redundant labels and masks; The helmet information was bound to the head as the detection target and the Soft-NMS was introduced to reduce the interference of occlusion and handheld the helmet. Through the analysis of 500 evaluation data sets, the results yield the Mean Average Precision (mAP) of the algorithm of 97.1%, the Precision (P) of 98.0%, and the Recall (R) of 97.4%, which can meet the requirements of intelligent monitoring in different scenarios of power plants.