Fine-Grained PCB Defect Detection Transformer Based on Feature Enhancement with Defect-Free Samples

This repository is the official implementation of Fine-DETR, which is built upon the RTDETR codebase. Fine-DETR focuses on enhancing defect detection performance by mining pattern information from defect-free samples, addressing the limitation that traditional methods ignore the value of defect-free samples.

📌 Main Contributions & Methods

Core Modules of Fine-DETR

1. Semantic Background Suppression (SBS) SBS separates defect features and defect-free features, and adopts different learning strategies for each, ensuring the stable learning of the model on defect-related features.
2. Intersection Box Guided Perception (IBGP) IBGP strengthens the learning of inconspicuous features, further enhancing the model's ability to learn defect features.

Extended Implemented Modules

To further optimize detection performance, the repository also implements the following modules:

• Loss functions: CIoU, EIoU, alpha-IoU
• Detection head: cosine detection head
• Attention mechanism: CBAM residual attention mechanism
• Downsampling method: Adown efficient downsampling method

🙏 Acknowledgements

This repository is built based on the RTDETR codebase. We sincerely thank the authors of RTDETR for their open-source contribution.