Abstract

This paper presents a generative deep learning approach to mitigate image degradation caused by partial coherence in lensless Coherent Diffractive Imaging (CDI). While high-resolution CDI relies on ideal X-ray coherence, experimental constraints frequently introduce partial coherence, leading to blunted diffraction patterns. Unlike existing methods that require expensive computation or pre-calibration to model illumination coherence, our proposed Generative Adversarial Network (GAN) directly restores sharp structural features from degraded images without explicit coherence characterization. The effectiveness of this approach is validated through applications in both CDI and ptychography, indicating its potential utility for a wide spectrum of phase-retrieval-based imaging techniques.