Improved YOLO Model Detects Tiny PCB Defects While Cutting Model Size

Printed circuit boards are among the most scrutinized components in electronics manufacturing, yet their surface defects remain notoriously hard to catch automatically. Flaws such as gold-surface discoloration, soldering residues, foreign objects and miscellaneous blemishes are typically small, low-contrast and easily lost against complex circuit backgrounds. A new study published in Cluster Computing reports a reworked YOLO-based detection model that improves sensitivity and specificity on such defects while shrinking the network for lightweight deployment.

The researchers first assembled an industrial PCB defect dataset covering the four defect categories. On top of a mature YOLO framework, they designed a GSAM attention block that integrates the GAM and SimAM attention mechanisms into the C3 module, with the aim of strengthening feature responses for small defects. They also introduced an improved mBiFPN structure into the network neck to reinforce fusion between shallow and deep multi-scale features.

The intermediate model built from these two changes, named YOLOv5-GB, raised sensitivity by 2.54 percentage points and specificity by 6.16 percentage points relative to the original YOLO model under the same test conditions, according to the abstract.

To cut computational cost further, the team swapped in ShuffleNet v2 as a lightweight backbone. Direct backbone replacement reduced the parameter count to 57.16% of the original model and improved detection speed by 3.80%, but it came at a price: sensitivity dropped by 8.54 percentage points. The final Improved-YOLO model therefore combines GSAM, mBiFPN and ShuffleNet v2 to balance small-defect detection accuracy against lightweight deployment requirements.

In the reported experiments, Improved-YOLO achieved a sensitivity of 90.63% and an F-score of 87.09%, while reducing FLOPs, MACs, parameter count and model size compared with YOLOv5-GB. The authors also included YOLOv12l as a recent external reference from the YOLO series, to provide a contemporary engineering comparison of model complexity and detection performance.

The work addresses a practical tension in industrial inspection: attention modules and multi-scale fusion tend to improve the detection of small, easily confused defects, but they add computation, while lightweight backbones cut cost at the expense of accuracy. By reporting sensitivity, specificity, F-score, speed and model size together, the study frames the trade-off explicitly rather than optimizing a single metric. The findings are based on the authors’ own PCB defect dataset and reported benchmark comparisons, so performance on other defect types or production lines would need separate validation.

Subject of Research: A lightweight improved YOLO deep learning model for detecting surface defects on printed circuit boards

Article Title: A lightweight and accurate improved YOLO model for PCB surface defect detection

Article References: Zhang, J., Liu, Y., Xu, H., & Tang, H. (2026). A lightweight and accurate improved YOLO model for PCB surface defect detection. Cluster Computing, 29(16), Article 917. https://doi.org/10.1007/s10586-026-06665-x

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06665-x

Keywords: printed circuit boards, defect detection, YOLO, deep learning, attention mechanism, ShuffleNet v2, lightweight models, computer vision, industrial inspection, feature fusion, lightweight, accurate