How Can a Super-Lightweight Algorithm Revolutionize Power Equipment Detection?

During power‑line inspections, spotting transformers, insulators and other components is crucial to prevent blackouts. So far, computer‑vision systems used for this task were bulky and demanded huge computational power, making them hard to deploy on lightweight drones or robots. The paper “LRA‑YOLO” offers a solution: a detection algorithm that merges three key ideas. First, the CRepLK module builds a huge receptive field using large‑kernel convolutions, but without inflating the model, thanks to a re‑parameterization technique that merges multiple branches into a single operation. Second, the LEDSF “Neck”, guided by ELA attention, better aligns extracted features, avoiding the misalignment errors typical of traditional architectures. Finally, the PADH head exploits redundancy in activation maps to drastically cut the number of parameters. The results are striking: compared to the baseline model, parameters, GFLOPs and file size shrink by almost half, while precision (mAP) rises by 1 %. Tests on three different datasets confirm that LRA‑YOLO balances speed and accuracy, paving the way for more agile drones and more frequent inspections. In practice, it’s like giving a tiny robot the eyesight of an elephant, without weighing it down.

The core of LRA‑YOLO is the CRepLK (Convolutional Re‑parameterized Large Kernel) module. In practice, the network builds multiple convolutional branches with different kernel sizes (3×3, 5×5, 7×7) and merges them during training. At inference time, these branches are fused into a single large‑kernel convolution, reducing the number of operations and memory footprint. This strategy enhances the model's ability to capture patterns at various scales, essential for recognizing objects of varying sizes along power lines. The LEDSF (Lightweight ELA‑guided Dynamic Sampling Fusion) Neck introduces an attention mechanism based on Edge‑Level Attention (ELA). ELA assesses the importance of edges in each feature map, guiding a dynamic fusion that better aligns information from different network levels. This mitigates the feature misalignment issue typical of traditional FPNs (Feature Pyramid Networks). The PADH (Partial Asymmetric Detection Head) exploits redundancy in activation maps: instead of processing all maps with symmetric convolutions, it uses only a subset, applying asymmetric kernels (e.g., 1×3 and 3×1) to capture specific directions. The result is a lighter detection head with fewer parameters while preserving precision. The entire architecture was evaluated on the Insplad dataset, achieving a 46.8 % reduction in parameters, 44.1 % fewer GFLOPs, and a 46.4 % decrease in model size, with a 1 % mAP increase over the baseline. Tests on two additional power equipment datasets confirm the design's robustness, showing that LRA‑YOLO maintains an optimal balance between computational efficiency and detection accuracy, making it ideal for resource‑constrained platforms such as drones or autonomous robots.