Research on Lightweight Edge Power Equipment Defect Recognition Model Integrated with Attention Mechanism

Aiming at the challenges in power equipment inspection images, including tiny defective targets, complex backgrounds, limited deployment conditions of edge devices, and the difficulty of existing models balancing detection accuracy and inference efficiency, this paper proposes a lightweight defect recognition model LAFD-Net integrated with attention mechanism for edge power equipment. The backbone network adopts depthwise separable convolution, Ghost feature generation and lightweight residual connection to reduce model parameters and computational complexity. A multi-attention fusion module is designed to enhance the response of critical defect areas from channel, spatial and positional dimensions. A multi-scale and multi-level feature fusion structure is constructed to improve the detection performance of small targets and weak-texture defects. Meanwhile, an adaptive keypoint detection head is introduced to boost the localization accuracy of defects with ambiguous boundaries. Experimental results show that LAFD-Net achieves a parameter volume of 7.54 M and computational complexity of 18.2 G. Its mAP@0.5 and mAP@0.5:0.95 reach 94.3% and 63.2% respectively, with an inference speed of 118 FPS on the server platform. Deployed on Jetson Xavier NX edge device, the single-frame inference latency is 18.3 ms, the detection speed is 54.6 FPS and power consumption is 12.2 W. The proposed model realizes an optimal trade-off among detection accuracy, model size and edge inference efficiency, and can satisfy the real-time defect recognition requirements of power inspection.