A multi-cognitive PCB defect detection model integrating Mamba.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Printed Circuit Board (PCB) defect detection faces challenges from complex textures, small targets, and subtle class similarities.
• Traditional methods lack robustness; deep learning models struggle with feature extraction, utilization, and complexity.
• Existing detectors often fail to efficiently extract tiny defect features or manage model complexity.

Purpose of the Study:

• To develop a robust and efficient framework for high-precision PCB defect detection.
• To address limitations in current deep learning models for detecting small and complex defects on PCBs.
• To introduce a hybrid approach combining Convolutional Neural Networks (CNNs) and Mamba state space models.

Main Methods:

• Proposed PCB-MMF framework with an MM-NET backbone integrating CNNs and Mamba for local and global feature modeling.
• Implemented a Three-Stage Multi-Receptive Module (TSMR) for feature fusion and redundancy reduction.
• Introduced a Multi-Cognitive Visual Augmentation Module (MC-VAM) for enhanced attention and shallow feature preservation, and a Lightweight Group-Shared Detection Head (LGSD) for computational efficiency.

Main Results:

• PCB-MMF achieved high mAP50 scores: 93.43% (HRIPCB), 98.68% (DeepPCB), and 85.39% (DsPCBSD+).
• Demonstrated robust generalization on the NEU-DET dataset with an mAP50 of 76.69%.
• Significantly reduced parameters and computational load compared to YOLO11 (8.9% fewer parameters, 12.70% less FLOPs) and Mamba-YOLO (58.48% fewer parameters, 55.28% less FLOPs).

Conclusions:

• PCB-MMF offers a superior balance of accuracy and efficiency for industrial PCB defect detection.
• The hybrid CNN-Mamba approach effectively addresses challenges in detecting small and complex defects.
• The proposed framework provides a promising solution for lightweight, high-precision automated optical inspection systems.