YOLO-UMS: Multi-Scale Feature Fusion Based on YOLO Detector for PCB Surface Defect Detection

Area of Science:

• Electronics manufacturing
• Computer vision
• Artificial intelligence

Background:

• Printed circuit board (PCB) defect detection is crucial for electronics quality control.
• Complex PCB layouts present challenges for automated inspection, including low contrast, uneven brightness, and small, irregular defects.
• Existing methods struggle with rapid and accurate detection of these defects.

Purpose of the Study:

• To propose a novel object detector, YOLO-UMS, for enhanced accuracy and speed in PCB surface defect detection.
• To address limitations in detecting defects with low image contrast, uneven brightness, small sizes, and irregular shapes.
• To provide an efficient and accurate solution for industrial PCB inspection.

Main Methods:

• Developed a lightweight Unified Multi-Scale Feature Fusion Pyramid Network (UMSFPN) for effective multi-scale information fusion.
• Introduced a lightweight RG-ELAN module to improve feature extraction for small targets in complex scenes.
• Integrated an Adaptive Interaction Feature Integration (AIFI) module and a channel-priority convolutional attention module (CPCA) to enhance feature representation.

Main Results:

• The UMSFPN neck improved AP50 by 3.1% and AP by 2% on the PCB-M dataset compared to the original PAFPN.
• YOLO-UMS achieved an 84% AP50 on the PCB-M dataset without pre-training, a 6.4% improvement over the YOLO11 baseline.
• The proposed method demonstrated strong detection accuracy compared to existing algorithms, showing broad applicability across datasets and detectors.

Conclusions:

• YOLO-UMS offers a feasible and effective solution for efficient and accurate PCB surface defect detection.
• The novel UMSFPN, RG-ELAN, AIFI, and CPCA modules contribute to improved performance in challenging inspection scenarios.
• The algorithm shows significant potential for industrial application in automated PCB quality control.