LXIE-Net and HLXray: A Mamba-Based Network and Real-World Dataset for Low-Dose X-Ray Image Enhancement in Industrial Inspection

In industrial inspection, achieving clear X-ray imaging requires a high radiation dose, which poses risks such as damaging inspected targets, impacting operator health, and reducing inspection efficiency due to longer imaging times. Thus, ensuring inspection accuracy while minimizing radiation dose has become an urgent practical requirement. An effective way is using image enhancement technology to improve the quality of low-dose images, achieving the same clarity as high-dose images. However, existing methods, such as those using Transformer backbones, struggle to balance capturing global information (critical for accurate industrial inspection) with reducing computational resource consumption. On the other hand, existing datasets lack paired low- and high-dose images, both from real-world sources, for training these models. For instance, some low-dose images are often generated by artificially adding noise to real high-dose images. As a result, although methods achieve high accuracy on these datasets, applying them to real-world industrial inspections remains challenging. To solve these issues existing in both method and dataset, in this paper, we propose a Low-dose X-ray Image Enhancement network (LXIE-net) based on Mamba architecture ar for industrial inspection, and a large-scale dataset with real-captured paired high- and low-dose X-ray images, named HLXray. To further demonstrate the effectiveness of our method in real-world industrial inspection scenarios, we annotated 800 real images with three types of defects in the HLXray dataset. In addition, we simulated low-dose images on the public GDXray dataset by adding noise to the corresponding high-dose images. Based on these two datasets, we conducted industrial defect detection experiments using low-dose images, generated high-dose images (i.e., enhanced low-dose images produced by our method), and real high-dose images for comparison. The results show that using the generated high-dose images significantly improves detection accuracy compared to using the original low-dose images, and the performance is nearly consistent with that of the real high-dose images. It demonstrates the effectiveness of applying our method in real industrial inspection application. The code and dataset are publicly available at https://github.com/YqunYang/LXIE-net.