Abstract:

In UAV power inspection，due to complex backgrounds，varying lighting conditions，and tiny sizes of insulator defects，it is difficult to balance the detection accuracy and efficiency of the inspection. To address the problem，ADC-YOLO，a lightweight dynamic attention detector based on YOLOv8n was proposed. The core of the model lay in the design of an Attentive Dynamic Convolution（ADC）module that serially connected lightweight dynamic convolution with the Coordinate Attention（CA）mechanism within a unified computational unit，constructing a synergistic feature extraction paradigm that integrated content adaptation and spatial refinement. By embedding the ADC module into the backbone structure of YOLOv8n，the model's feature extraction and multi-scale fusion capabilities were significantly enhanced. Experimental results showed that compared to mainstream lightweight detectors such as YOLOv8n，ADC-YOLO maintained lower levels of computational complexity and parameter count on the self-built High-Resolution Insulator Defect Dataset（HR-IDD）. ADC-YOLO achieved 0.806 and 0.445 in the mAP@0.5 and mAP@0.5∶0.95 metrics，respectively，outperforming all comparative models. Compared to the baseline model's mAP@0.5（0.790）and mAP@0.5∶0.95（0.435），the improvements were 2.0% and 2.2%，respectively，meeting the requirements for insulator defect detection.

Key words: insulator defect detection, lightweight network, dynamic convolution, coordinate attention, object detection, YOLO, UAV power inspection