"PI-MagNet-datasets"

"This is the translated English version of your README.md, with the \"Environment Setup\" section removed as requested. --- # PI-MagNet: Physics-Informed Lightweight Attention Network for High-Accuracy MFL Welding Defect Detection This is the official implementation of the paper: **\"Physics-Informed Lightweight Attention Network for High-Accuracy Magnetic Flux Leakage Welding Defect Detection.\"** PI-MagNet is a lightweight network specifically designed for industrial welding defect detection. By integrating Magnetic Perturbation Gradient (MPG) and Magnetic Perturbation-Aware (MPA) attention mechanisms, it overcomes the \"physics-blind\" limitations of conventional deep learning models. It achieves superior accuracy (92.1% mAP@0.5) with a highly compact footprint of only 2.50M parameters. ## \u2728 Key Features - **Physics-Informed**: Integrates deterministic electromagnetic principles (Gauss\u2019s law for magnetism and Magnetic Dipole Model) into the network architecture.- **MPG Enhancement Module**: Extracts 2D spatial gradients to amplify defect-induced structural variations while suppressing low-frequency \"magnetic pedestal\" background noise.- **MPA Spatial Attention**: Uses gradient-derived cues to steer the network's focus toward defect-sensitive regions based on field-consistency violations.- **Lightweight Design**: Incorporates **GhostConv** to eliminate feature redundancy, reducing parameters by over 94% compared to YOLOv8, making it ideal for edge devices like NVIDIA Jetson Nano.- **Physics-Constrained Loss ($L_{phys}$)**: Implements Magnetic Perturbation Divergence (MPD) and Magnetic Leakage Source (MLS) constraints to anchor predicted bounding boxes to physical leakage peaks and prevent \"box drift.\" ## \ud83d\udee0 Architecture Overview   *(Recommended: Insert Fig. 2 from your paper here)* ## \ud83d\udcca Open-Source Dataset (HC430LA MFL Dataset) This project releases the high-resolution 2D Magnetic Flux Leakage (MFL) image dataset used in the study:- **Hardware**: Captured by a customized flexible 32\u00d732 z-axial Hall sensor array.- **Resolution**: Super-resolution enhancement yielding ~0.5 mm per pixel.- **Quantity**: 1,000 high-resolution MFL images (800 training \/ 200 validation).- **Material**: Automotive-grade HC430LA high-strength steel welds.- **Download**: [Link to Zenodo \/ IEEE DataPort \/ Hugging Face] ## \ud83d\ude80 Quick Start ### 1. Data PreparationOrganize your dataset in the following structure:```text\/data  \/images    \/train    \/val  \/labels    \/train    \/val``` ## \ud83d\udcc8 Performance Comparison Benchmark results on the automotive-grade HC430LA steel weld dataset: | Model | mAP@0.5 (%) | Parameters (M) | Deployment || :--- | :---: | :---: | :--- || YOLOv8 | 87.6 | 43.61 | GPU || YOLOv11 | 80.7 | 25.28 | GPU || FasterNet | 81.9 | 3.85 | Edge\/GPU || EfficientViT | 86.9 | 41.50 | GPU || **PI-MagNet (Ours)** | **92.1** | **2.50** | **Jetson Nano (18 FPS)** | Here is the updated **Citation** section and a **Project Status** notice for your README.md. I have modified the BibTeX to reflect the \"Under Review\" status as requested. --- # PI-MagNet: Physics-Informed Lightweight Attention Network for High-Accuracy MFL Welding Defect Detection > **Note:** This repository corresponds to a research paper currently **Under Review**. Full citation details and the official DOI will be updated upon publication. ... (Keep the rest of your sections like Key Features, Architecture, etc.) ... ## \ud83d\udd17 Citation If you use this code or our imaging method in your research, please cite the following papers.  ### 1. PI-MagNet Algorithm (Main Paper - Under Review)```bibtex@article{shi2026physics,  title={Physics-Informed Lightweight Attention Network for High-Accuracy Magnetic Flux Leakage Welding Defect Detection},  author={Shi, Chengyu and Li, Xuan and Zhang, Yi and Zhu, Hongxi and Li, Chu and Ji, Guanghao and Huang, Yunkai and Sun, Siyi and Pei, Zhengyang and Zhang, Mingji},  journal={Under Review},  year={2026},  note={To be updated upon publication}}``` ### 2. MFL Imaging Foundation (Hardware & Data Foundation)```bibtex ``` ## \ud83d\udce7 Contact "