Research on deep learning restoration algorithm of X-ray backscatter imaging based on virtual training dataset.

Area of Science:

• X-ray optics and imaging
• Deep learning for image processing
• Scientific instrumentation

Background:

• X-ray lobster eye lenses offer wide-field imaging but suffer from image degradation due to their point spread function (PSF).
• Existing image restoration methods are insufficient for effectively removing noise and artifacts introduced by the PSF.
• Limited availability of real-world backscatter datasets hinders the development of robust restoration models.

Purpose of the Study:

• To develop an effective image restoration technique for X-ray lobster eye lens imaging.
• To overcome the limitations of sparse real-world datasets by utilizing a virtual training dataset.
• To improve the signal-to-noise ratio and structural fidelity of backscatter X-ray images.

Main Methods:

• A virtual training dataset was generated by convolving the PSF with object simulations to mimic image degradation.
• This synthetic dataset was used to train a deep learning model for image restoration, eliminating manual annotation.
• The trained model was then applied to restore actual backscatter X-ray images.

Main Results:

• The virtual training dataset approach successfully restored real backscatter images with high accuracy.
• Quantitative evaluation showed a structural similarity index (SSIM) of 0.86 and a mean intersection over union (MIoU) of 0.83.
• The method effectively mitigated noise and degradation caused by the X-ray lobster eye lens PSF.

Conclusions:

• The proposed virtual training dataset method offers a viable solution for restoring degraded X-ray lobster eye lens images.
• This approach significantly reduces reliance on scarce real-world data, saving time and resources.
• The technique enhances system safety by enabling reduced radiation flux and shorter acquisition times.