Cable Bracket Tilt Detection Based on YOLOv3

Aiming at the problems of large number of brackets, small targets and overlapping perspectives in bracket tilt detection in cable tunnels, an improved YOLOv3 algorithm is proposed. Optimize the Darknet-53 feature extraction network; improve the target detection layer; weight the center coordinate loss and confidence loss according to the location information; use the K-Means clustering algorithm to eliminate interference data. The experimental results show that, compared with the traditional YOLOv3 algorithm, the average accuracy of cable bracket recognition has increased from 82.34% to 86.73%, the detection frame rate has increased from 47 FPS to 54 FPS, and the cable bracket tilt recognition accuracy has reached 92%.