Intelligent Defect Recognition Method Based on Multi-Scale Feature Fusion in UAV Inspection of Transmission Lines

Abstract

To address the problems of large defect scale differences, easy loss of texture of small targets, strong interference from complex backgrounds, and limited computing power at the edge in UAV inspection images of transmission lines, a multi-scale feature fusion detection method, MSFF-TLDet, is proposed. This method sets up a high-resolution branch and a non-negative weighted bidirectional fusion network, and combines cross-scale background suppression and decoupled detection heads to enhance the response of small targets at different shooting distances. To verify the independent contribution of each module, a parameterized simulation environment is established based on the category ratio, target scale, and imaging degradation law of publicly available inspection data, and four types of experiments are set up: model comparison, training convergence, comprehensive perturbation, and module ablation. Results show that MSFF-TLDet's Precision, Recall, mAP@0.5, and APsmall are 0.925, 0.904, 0.929, and 0.613, respectively, which are 3.2, 4.3, 3.2, and 7.5 percentage points higher than YOLOv8s; the model has 9.6 M parameters and an inference speed of 69 frames/s. When the disturbance intensity reaches 40%, the mAP@0.5 is still 0.782, indicating that adaptive fusion can alleviate the feature degradation caused by resolution decrease, blurring and occlusion.