YOLO-APGC: A lightweight method for BGA surface defect detection based on cross-scale feature fusion

Abstract

Ball Grid Array (BGA) surface defect detection is critical for ensuring the safety and reliability of electronic products. In recent years, numerous deep learning-based BGA surface defect detection algorithms have achieved significant detection performance in this field. However, the computational complexity of existing high-precision models has limited their application in real-time detection on industrial production lines. To address this issue, this paper proposes a lightweight detection architecture based on You Only Look Once v8 (YOLOv8), named YOLO-ADown and Partial Group-shuffle Cross-scale Feature Fusion (YOLO-APGC), which aims to significantly reduce the number of model parameters and computational costs while maintaining high detection accuracy. First, in the feature extraction stage, Adaptive Downsampling (ADown) is introduced, and a novel lightweight Double Partial-Block (PP-Block) module is constructed to synergistically optimize the information retention efficiency and computational cost during the high-dimensional feature compression process. Then, to achieve efficient fusion of multi-scale features, a novel cross-scale dynamic feature fusion network, the Partial Group-shuffle Cross-scale Feature Fusion Network (PG-CCFN), is proposed. Finally, a new BGA dataset is constructed based on industrial scenarios, and a series of experiments are conducted. The results show that the number of parameters in YOLO-APGC is reduced by 68.21 % compared to the baseline model. Additionally, the Giga Floating-point Operations Per Second (GFLOPs) is reduced by 54.32 %, and the Frames Per Second (FPS) is improved by 9.1 %. This model offers a low-cost, highly robust intelligent quality inspection solution for the microelectronics packaging industry, providing significant engineering value in ensuring the long-term reliability of microelectronics packaging.