LAC-NET: robust lightweight adaptive convolutional network for insulator defect detection within complex scenarios

Reliable insulator defect detection of transmission lines is essential for ensuring the safe operation of power systems. However, existing detection models still suffer from insufficient discriminative feature extraction and inaccurate defect localization in complex backgrounds. Lightweight models are often affected by the diversity of defect appearances, whereas high-accuracy models typically incur substantial computational costs. Achieving an appropriate balance between detection accuracy and computational efficiency therefore remains a critical challenge. To address this issue, a task-oriented Lightweight Adaptive Convolutional Network (LAC-Net) is proposed for insulator defect detection, specifically considering the small scale, irregular morphology, and complex backgrounds of defect targets. First, Linear Deformable Convolution (LDConv) is incorporated into selected feature-extraction layers, allowing the convolutional sampling pattern to adapt to irregular and deformed defect regions and thereby improving feature representation and defect localization. Second, the Dynamic Up-sampler (DySample) is introduced into the feature-fusion network to replace conventional upsampling, enabling efficient dynamic feature reconstruction with reduced computational overhead. Finally, the LAC-Net architecture is established to achieve a favourable trade-off between detection accuracy and computational complexity in complex inspection scenarios. Experimental results demonstrate that LAC-Net improves mAP@0.5 by 5.9 percentage points over the baseline YOLOv11s, while providing faster inference and lower computational cost.