Faster R-CNN Transmission Line Multi-target Detection Based on BAM

Abstract

Aiming at the problems that the existing aerial inspection algorithm in the transmission line is affected by the environment and complex background and other factors, resulting in poor detection effect, low accuracy and poor real-time performance. A recognition algorithm for multi-target detection of transmission lines based on BAM's Faster R-CNN is proposed. Firstly, the structure of the traditional Fast-R-CNN feature extraction network is optimized by using the residual network ResNet50 as its backbone feature extraction network and adding BAM (bottleneck attention module) to the network to improve the visibility and accuracy of the target region in the image. Secondly, we use Softer NMS instead of NMS to improve the non-maximal suppression of the images taken during the inspection. Then combined with the K-means++ clustering algorithm to optimize the anchor parameters. Finally, the FPN feature pyramid network is integrated to further improve the detection accuracy of the algorithm. The experimental results show that the detection speed and accuracy of the improved Faster R-CNN have been effectively improved, achieving a gain of +4.3 mAP.