Towards High Quality PCB Defect Detection Leveraging State-of-the-Art Hybrid Models

Abstract

The automatic detection of defects in printed circuit boards (PCBs) is a critical step in ensuring the reliability of electronic devices. This paper introduces a novel approach for PCB defect detection. It incorporates a state-of-the-art hybrid architecture that leverages both convolutional neural networks (CNNs) and transformer-based models. Our model comprises three main components: a Backbone for feature extraction, a Neck for feature map refinement, and a Head for defect prediction. The Backbone utilizes ResNet and Bottleneck Transformer blocks, which are proficient at highlighting small defect features and overcoming the shortcomings of previous models. The Neck module, designed with Ghost Convolution, refines feature maps. It reduces computational demands while preserving the quality of feature representation. This module also facilitates the integration of multi-scale features, essential for accurately detecting a wide range of defect sizes. The Head employs a Fully Convolutional One-stage detection approach, allowing for the prediction process to proceed without reliance on predefined anchors. Within the Head, we incorporate the Wise-IoU loss to refine bounding box regression. This optimizes the model's focus on high-overlap regions and mitigates the influence of outlier samples. Comprehensive experiments on standard PCB datasets validate the effectiveness of our proposed method. The results show significant improvements over existing techniques, particularly in the detection of small and subtle defects.