RefDiffNet: Learning to Expose Subtle PCB Defects Before Detection

RefDiffNet introduces a lightweight neural network module that enhances PCB defect detection by comparing defective images against reference images, improving detection accuracy by up to 18% while adding minimal computational overhead. The plug-and-play approach works across multiple detector architectures, bridging classical inspection techniques with modern deep learning.

RefDiffNet addresses a critical challenge in manufacturing quality control by reviving a classical inspection principle—reference-based comparison—within contemporary deep learning frameworks. PCB defect detection has historically struggled with false negatives on subtle defects obscured by complex background patterns, a problem that persists even with state-of-the-art neural detectors relying solely on the defective image itself.

The innovation stems from recognizing that defect-free reference images contain valuable structural information that standalone detection misses. By computing differences between aligned reference and defective images, RefDiffNet isolates anomalies and pre-processes them into a form that downstream detectors can more easily interpret. This input enhancement approach is elegant: it requires only 0.004-0.005M additional parameters and 0.7-0.8 GFLOPs, representing less than 0.25% computational overhead while delivering 18% relative improvement in mAP50:95 metrics.

The detector-agnostic nature of RefDiffNet—demonstrated across YOLOv8 through v26, RT-DETR, and Faster R-CNN—signals practical utility in production environments where detector architectures evolve. For manufacturers, this translates to improved defect catch rates without expensive hardware upgrades, directly impacting quality assurance costs and yield rates. The research validates that hybrid approaches combining classical inspection wisdom with modern deep learning architectures outperform pure learning-based methods, particularly in domains where reference standards exist.

Future developments may explore temporal reference integration in manufacturing lines or extension to other industrial inspection domains sharing similar reference-based characteristics, suggesting broader applications beyond PCBs.

• →RefDiffNet achieves up to 18% mAP improvement on PCB defect detection with negligible computational cost (0.004-0.005M parameters)
• →The approach combines classical reference-based inspection with deep learning, making subtle defects more visible to downstream detectors
• →Works as a plug-and-play enhancement module compatible with multiple detector architectures from one-stage to transformer-based models
• →Adds less than 0.25% parameter overhead to any evaluated detector while significantly improving detection accuracy
• →Demonstrates that hybrid classical-modern approaches outperform pure learning methods in domains with reference standards