Deformable convolution and multi-scale fusion for high-performance transmission line defect detection

This article presents a high-performance defect detection algorithm for unmanned aerial vehicle (UAV) transmission line inspection, addressing issues such as missed detections, false alarms, and low accuracy. The algorithm combines the C3K2-DCNv4 module, which adopts a hierarchical integration strategy and which leverages deformable convolution to better capture geometric features from irregular targets. It also integrates the SPPF-LSKA module with a large kernel separation convolution attention mechanism, replacing the original SPPF structure to improve multi-scale feature extraction while reducing model redundancy. The neck network is redesigned with a diffuse pyramid structure integrated with DASI to enhance multi-scale feature fusion and context propagation. Additionally, the DyHead module incorporates a dynamic attention mechanism to strengthen feature representation under complex backgrounds. The introduction of the Focal-DIoU loss function and a focus on difficult samples, along with bounding box regression optimization, enhance convergence speed and localization accuracy. The main contribution of this work is the task-driven configuration and coordinated interaction of these modules for UAV-based transmission-line inspection. Specifically, DCNv4 is selectively inserted into the P2-P4 backbone stages, the P2 detection branch is added while the P5 branch is removed to prioritize small and medium-sized defects, SPPF-LSKA is channel-reduced to control computational redundancy, and the diffuse pyramid, DASI, DyHead, and Focal-DIoU are jointly configured to address cross-scale information loss, complex backgrounds, and hard imbalanced samples. Experimental results demonstrate that the proposed model outperforms the baseline, with an 8.4-percentage-point increase in [email protected], a 4.8-percentage-point increase in precision, and an 11.8-percentage-point increase in recall. On an NVIDIA Jetson Orin NX 16GB platform, the measured TensorRT INT8 latency was 128 ms for a batch of 16 images, corresponding to an effective throughput of 125 images/s. Under FP16 precision, the measured latency was 144 ms per batch, corresponding to 111 images/s. These results demonstrate efficient batched inference under the tested configuration. Because the Jetson evaluation used a batch size of 16, the results should be interpreted as batched throughput rather than single-frame camera latency.