XPtychoSR: a self-supervised super-resolution framework for X-ray ptychography

Ptychography has been extensively used for nanoscale imaging of catalysts, cells, and chips. The interpolation-based magnification is no longer adequate. A clearer imaging quality is expected for observing fine structures in the reconstructed images. While deep-learning-driven super-resolution (SR) techniques have been successful, pre-trained or supervised strategies are still challenging in X-ray ptychography experiments as generalization and large training datasets need to be considered. A self-supervised SR method named XPtychoSR is proposed to further discern the reconstruction details in the large field of view (FOV) X-ray ptychography data. XPtychoSR generates clear SR results based on only one input, without pre-training and matched (or unmatched) datasets. XPtychoSR extracts the content prior and the edge prior of the input, which is fused with the implicit image prior constructed by a U-shaped net to enhance detail expressiveness in the SR image. A physics diffraction model (PDM)-based reconstruction method is used to guide the network to self-supervised learning, resulting in an SR image with superior detail. The simulation and soft X-ray ptychography experiment show that XPtychoSR has better detail resolution ability than the existing advanced SR technologies. Ablation experiments demonstrate the effectiveness of the components and framework in XPtychoSR. XPtychoSR adapts to the need for SR imaging in the large FOV and detailed viewing in the region of interest (ROI). The method avoids additional adjustments and costs in hardware. These improvements further extend the application of X-ray ptychography.