Restoration of X-ray backscatter images under low exposure time based on deep learning

Lobster-eye lenses are suitable for staring hard X-ray backscattering imaging, but achieving clear imaging typically requires a long exposure time. This study establishes a noise model for imaging under low-exposure conditions, investigates the causes of image degradation, and develops a method combining variance-stabilizing transformation and convolutional neural networks to simplify the noise model and achieve denoising. The proposed method uses paired low-exposure data for training without the need for ground-truth images. Compared to traditional non-local means filtering and the original noise2noise method, the proposed approach significantly improves denoising performance. Furthermore, compared to two other lobster-eye backscattering image restoration algorithms, the proposed method achieves an average SSIM improvement from 0.1798 to 0.8473 for the multiple target images. This technique not only enhances the signal-to-noise ratio of X-ray backscattering images, reduces exposure time to less than 20 seconds, and lowers radiation doses to improve imaging safety but also preserves target details and brightness changes. It provides an effective solution for low-exposure imaging in fields such as medical diagnostics and aviation.