Pre-classification-learning-based monochromatization strategy for broadband coherent diffraction imaging

Coherent diffraction imaging (CDI) is a lensless imaging technique with significant applications across many fields. It utilizes coherent light to illuminate objects and achieves high-resolution reconstruction from the captured diffraction patterns. However, under broadband illumination, such as an attosecond pulse, the applicability and reconstruction quality of CDI are severely limited by low temporal coherence. Here, we developed a pre-classification learning-based monochromatization strategy for broadband CDI. Specifically, we employ a classification neural network to classify diffraction patterns with unknown bandwidths in advance, thereby enabling more effective monochromatization and accurate reconstruction. The proposed method demonstrates strong robustness across diverse spectral conditions and object types. Notably, it can achieve super-broadband CDI for both continuous and discrete spectra, even in the absence of spectral prior information. This method is expected to advance the development of learning-based broadband optical imaging systems.