YOLOv7-TID: A Lightweight Network for PCB Intelligent Detection

Abstract

The Printed Circuit Board (PCB), often regarded as the cornerstone of the electronic information industry, plays a pivotal role in determining the quality of electronic products. However, traditional detection methods struggle to identify minuscule defects on PCBs due to their small surface area and intricate layouts, leading to a decline in product quality. To address these challenges, we propose an innovative lightweight intelligent detection algorithm, named YOLOv7-tiny Improved Detection (YOLOv7-TID). For precise defect detection on PCBs. This new model, based on YOLOv7-tiny, incorporates several enhancements. First, a parallel network module is added to the neck network to bolster the backbone network’s ability to extract feature information from both deep and shallow layers of images. Second, the sampling scale for small targets is increased, and the FPN and PAN structures are improved to enhance the feature extraction network’s semantic feature extraction and localization capabilities. Additionally, the SimAM attention mechanism module should be introduced to improve the network’s focus on shallow features without increasing the number of parameters. The model is further optimized by using the slim-neck network and the DWConv convolution module to reduce its weight, and by employing the NWD loss function to calculate positioning loss and enhance the network’s detection capability for small targets. Finally, the lightweight model undergoes pruning and knowledge distillation. Experimental results show that, compared with the original YOLOv7-tiny, the new model achieves a detection accuracy of 96.4% on the PCB defect test dataset. The [email protected] and [email protected]:.95 are increased by 6.2% and 6.0%, respectively. Additionally, the number of parameters is reduced to only 3.5M, and the computation load is decreased by 23.5%. This makes the model more suitable for industrial applications and embedded development systems.

Keywords

• Defect detection
• deep learning
• NWD
• SimAM
• lightweight
• pruning