EHIR: Energy-based Hierarchical Iterative Image Registration for Accurate PCB Defect Detection

Paper

This repository is the official implementation of paper "EHIR: Energy-based Hierarchical Iterative Image Registration for Accurate PCB Defect Detection".

We propose EHIR (Energy-based Hierarchical Iterative Image Registration), a novel method that formulates image registration as an energy optimization problem using edge points instead of sparse features. This makes EHIR especially powerful in handling high-resolution images and those with scarce, dense, or degraded features, and significantly outperforms conventional feature-based methods.

Our method can be used for accurate PCB image registration and further reference-based defect detection (template image is required). Our framework consists of three stages: Edge-guided Energy Transformation (EET), Energy-based Hierarchical Iterative Image Registration (EHIR) and Edge-guided Energy-based Defect Detection (EEDD).

This approach can be extended to image registration problems characterized by prominent structural attributes.

Data Preparation

Please make sure that all template and target image pairs are placed together in a single directory (e.g., ./data). Each pair should be named in the following format: {name}_temp.jpg and {name}_test.jpg.

Usage

You can align your images by running the following script:

python tools/align.py

Acknowledgments

This work was contributed equally by Shuixin Deng and Lei Deng.

Citation

If you find our work useful, please consider to cite our work.

@article{deng2024ehir,
  title={EHIR: energy-based hierarchical iterative image registration for accurate PCB defect detection},
  author={Deng, Shuixin and Deng, Lei and Meng, Xiangze and Sun, Ting and Chen, Baohua and Chen, Zhixiang and Hu, Hao and Xie, Yusen and Yin, Hanxi and Yu, Shijie},
  journal={Pattern Recognition Letters},
  volume={185},
  pages={38--44},
  year={2024},
  publisher={Elsevier}
}