YOLOv8-DEE: a high-precision model for printed circuit board defect detection

Abstract

Defects in printed circuit boards (PCBs) occurring during the production process of consumer electronic products can have a substantial impact on product quality, compromising both stability and reliability. Despite considerable efforts in PCB defect inspection, current detection models struggle with accuracy due to complex backgrounds and multi-scale characteristics of PCB defects. This article introduces a novel network, YOLOv8-DSC-EMA-EIoU (YOLOv8-DEE), to address these challenges by enhancing the YOLOv8-L model. Firstly, an improved backbone network incorporating depthwise separable convolution (DSC) modules is designed to enhance the network’s ability to extract PCB defect features. Secondly, an efficient multi-scale attention (EMA) module is introduced in the network’s neck to improve contextual information interaction within complex PCB images. Lastly, the original complete intersection over union (CIoU) is replaced with efficient intersection over union (EIoU) to better highlight defect locations and accommodate varying sizes and aspect ratios, thereby enhancing detection accuracy. Experimental results show that YOLOv8-DEE achieves a mean average precision (mAP) of 97.5% and 98.7% on the HRIPCB and DeepPCB datasets, respectively, improving by 2.5% and 0.7% compared to YOLOv8-L. Additionally, YOLOv8-DEE outperforms other state-of-the-art methods in defect detection, demonstrating significant improvements in detecting small, medium, and large PCB defects.

Keywords

• Defects detection
• Computer vision
• PCB
• Deep learning
• YOLO