A Rapid Re-parameterized Convolutional Network for Detection of Electric Power Facilities

Fast and effective recognition of electric power facilities is critical for civilian and military applications of the digital electrical grid system. In this paper, we build a multi-scale dataset of electric facilities, and propose a new object detection method for various electric power facilities based on a parameter reconstructed Mask Region-based Convolutional Network (Mask RCNN). Our experiments are trained/tested on images of manually annotated electric power facilities, which are divided into two sections according to the scale of targets, including utility poles, transformers, insulators, cross arms, and wire clips. We utilize the Mean Shift (MS) algorithm to adapt the size of anchors in Region Proposal Network (RPN) for faster bounding region location. The experimental results demonstrate that compare to the origin model, our Rep-Mask RCNN can achieve 45.45% speedup in inference time when gets almost mean AP. For different backbone, our Rep-Mask RCNN can have 4.27% to 9.19% improvement for different size facilities.