EMR-net: A lightweight multi-scale feature fusion PCB defect detection algorithm

An efficient lightweight detection algorithm (EMR-net) is proposed to address the problems that the defects on the surface of PCB can be easily confused with the background and the resource consumption is too high due to the oversized model. First, an Edge Preservation model (EPstem) is proposed to efficiently extract the edge detail information of defects using the Sobel operator for the problem of insufficient image detail information extracted by ordinary convolution. Then, a Residual Multi-Scale Block (RMS Block) is designed to better preserve the detail features while reducing the spatial resolution of the feature map, and a Re-parameterization Convolution (RepConv) is introduced to efficiently extract defect features. Finally, a Context-Aware Routing for Pyramid Aggregation Network (CARPAN) is proposed in the neck part in combination with lightweight Channel Attention (CA) to effectively suppress irrelevant detail interference. Meanwhile, the Dynamic Sample (DySample) strategy is adopted to avoid the loss of target feature information that may be caused by the traditional up-sampling process. On the basis of keeping the model lightweight, the detection accuracy is significantly improved. The experimental results show that EMR-net achieves 93.5% in mAP index, which is 1.6% higher than YOLOv11s, and the number of parameters and computations are reduced by 46% and 33%, which demonstrates excellent practicability.