Defect Detection Model of Printed Circuit Board Components Based on the Fusion of Multi-Scale Features and Efficient Channel Attention Mechanism

The detection of defects in printed circuit board (PCB) components is crucial to the quality of PCB. Issues such as blurred details, complex and varied backgrounds, and inadequate recognition of PCB components lead to poor detection accuracy. To address these challenges, this paper introduces a PCB component defect detection model (MSF-ECANet) based on multi-scale features and efficient channel attention networks. Firstly, to address the challenge of unclear information regarding intricate features in deep networks, Residual Nets (ResNet) and Multi-Scale Feature Pyramid Networks (FPN) are integrated. This fusion tackles the issue of vanishing gradients, expands the model’s receptive field, and optimizes the model’s proficiency for recognizing PCB components. Secondly, to improve the recognition rate of PCB component detection, Efficient channel attention networks (ECA-Net) are used to assign different weights to the PCB background and foreground channels to segment the background and foreground. Lastly, a dichotomous K-means algorithm is used to obtain the optimal anchor size that is closer to the ground truth size, so as to improve the sensitivity of the model to small target detection. When compared to SSD, YOLOv3, YOLOv5, YOLOx and Faster R-CNN, the experimental results show that the model proposed in this paper improves 1.41%, 7%, 4.17%, 5.47% and 8.33% in accuracy, respectively. Furthermore, the improved network exhibits superior convergence compared to the original network. Therefore, the MSF-ECANet model presented in this paper is more suitable for industrial applications of PCB component defect detection.