Transmission line defect detection via an integrated improved YOLOv8 and deep neural random forest framework

Abstract

Abstract How to effectively identify defects in power transmission lines using unmanned aerial vehicle (UAV) technology remains a key research focus. This paper presents an optimization over traditional convolutional neural networks by proposing a spatially deformable convolution (SDC) algorithm to enhance feature extraction from images. Additionally, a more comprehensive hybrid loss function is introduced to improve the model's ability to recognize multi-scale defect patterns. A deep neural decision forest (DNDF) is then employed to perform fine-grained classification of candidate regions, outputting precise defect categories. Experimental validation shows that the proposed algorithm achieves recognition accuracy above 92% for five types of conditions: normal, stains, cracks, corrosion, and surface peeling. Compared with several conventional detection methods, the algorithm demonstrates notable improvements. For instance, relative to Faster R-CNN, it increases detection speed by 36 FPS and reduces the number of parameters by 54.57%. When compared with YOLOv7-M, YOLOv9-C, and YOLOv11-M, although inference speed is slightly reduced, the mean average precision (mAP) is improved by 2.6%, 2.38%, and 1.9%, respectively, due to increased algorithmic complexity. These results confirm that the proposed approach can effectively identify multi-scale defects in transmission lines.