Improved PCB Chip Recognition Method Based on YOLOv7-tiny

To address the inefficiency of manual visual inspection in PCB chip detection, this paper proposes an improved PCB chip recognition method based on YOLOv7-tiny, leveraging a machine vision platform and deep learning techniques. While YOLOv7-tiny is effective for PCB detection, it suffers from high miss rates, poor performance in small object detection, and deployment challenges. To overcome these issues, we introduce the lightweight convolutional neural network RepViT as the backbone, enhancing the network’s ability to capture fine-grained features while reducing deployment costs on resource-constrained mobile devices. Additionally, we replace the SE attention mechanism in RepViT with the more lightweight and efficient ECA module to reduce computational complexity, and incorporate DSConv to further decrease the computational load, thereby improving inference speed. Meanwhile, using a dataset of 3076 images from both publicly available and laboratory sources, the proposed method achieved an average precision of 97.2% for single-class PCB chip detection, representing a 0.8% improvement over YOLOv7-tiny. Moreover, despite a 15.9% increase in FLOPs, the model’s parameter count was reduced by approximately 45%, with significant enhancements in generalization performance compared to the original network.