A Lightweight YOLOv8n-Based Network with CAD and DSGE for Power Line Defect Detection

To address the sampling misalignment and detail loss caused by fixed-grid downsampling for small-scale defects, as well as the insufficient differentiated modeling and interaction of defect details and structural context in UAV-acquired power-line images, an enhanced lightweight YOLOv8n-based framework for power-line defect detection is developed. First, a content-adaptive downsampling (CAD) module is developed to predict input-dependent sampling offsets and normalized aggregation weights and to perform differentiable resampling. Combined with local-global interactive depthwise separable convolution, CAD improves the preservation of small-object details while maintaining relatively low computational complexity. Second, a dynamic subspace gated exchange (DSGE) module is proposed to adaptively partition features into a high-frequency detail subspace and a low-frequency structural subspace according to the input content. Heterogeneous branches and bidirectional gated exchange are then employed to jointly model fine-grained details and structural context. In addition, the lightweight mixed local channel attention (MLCA) mechanism is incorporated in the detection head as an auxiliary feature-enhancement component. Experimental results show that the proposed model achieves mAP@0.50 and mAP@0.50:0.95 values of 92.3% and 62.9%, respectively, outperforming the compared models under the current evaluation protocol. With 1.90 M parameters and 5.6 G FLOPs, the model reaches an inference speed of 134.7 FPS on the desktop GPU platform, demonstrating that content-adaptive sampling and dynamic detail–structure interaction can improve small-defect detection and complex-background suppression while maintaining relatively low model complexity.