基 于 混 合 深 度 先 验 的 快 照 式 宽 带 相 干 衍 射 超 快成 像 算 法(特 邀)

Objective Snapshot broadband coherent diffraction ultrafast imaging (BCDUI) is an emerging optical imaging paradigm that enables the capture of high-dimensional dynamic scenes with ultrahigh temporal resolution, large sequence depth, and rich physical information. By encoding temporal evolution into spectral measurements and reconstructing complex-amplitude data, this technique provides simultaneous access to both amplitude and phase, which are essential for revealing underlying physical processes such as morphology evolution, refractive index variation, and wavefront propagation. However, due to the intrinsic measurement mechanism that combines spectral multiplexing and spatial undersampling, the reconstruction of high-dimensional complex-amplitude data from two-dimensional measurements becomes a severely ill-posed inverse problem. This challenge is further exacerbated under low sampling rates and noisy conditions, where conventional iterative reconstruction algorithms often suffer from limited robustness, high computational cost, and insufficient reconstruction fidelity. Meanwhile, although data-driven deep learning approaches have demonstrated strong representation capabilities in inverse imaging problems, their performance heavily depends on large-scale training datasets, which are typically unavailable for ultrafast dynamic complex-amplitude imaging. In addition, pretrained models often exhibit limited generalization ability when applied to unseen physical scenarios. Therefore, there is a critical need for a reconstruction framework that can effectively integrate physical modeling and deep priors while avoiding reliance on extensive training data. To address these challenges, this work aims to develop a robust and high-fidelity reconstruction method for snapshot ultrafast complex-amplitude imaging under highly undersampled conditions. Methods In this study, a plug-and-play reconstruction algorithm based on hybrid deep priors (PnP-HDP) is proposed to solve the inverse problem in BCDUI. The method integrates a physics-driven deep image prior with a pretrained deep denoising network within alternating direction method of multiplier (ADMM) framework, enabling high-quality reconstruction without the need for large-scale training datasets. The forward imaging process is modeled as a mapping from high-dimensional complex-amplitude data to two-dimensional under-sampled intensity measurements, incorporating spectral-temporal encoding, coherent modulation, and spatial sampling. Based on this model, the reconstruction problem is formulated as a constrained optimization task consisting of a data fidelity term and a prior regularization term. To efficiently solve this problem, an augmented Lagrangian formulation is adopted, and the optimization is carried out using ADMM framework. A key feature of the proposed approach lies in the hybrid deep prior design. A dual-branch convolutional neural network is employed to represent the complex-amplitude data, where the amplitude and phase components are modeled separately using two parallel subnetworks. These subnetworks take fixed random noise as input and are optimized in a self-supervised manner driven by the physical forward model. This design leverages the structural bias of neural networks to implicitly regularize the solution space. In addition, a plug-and-play strategy is introduced by replacing the proximal operator in the iterative optimization with a pretrained deep denoiser. This allows the incorporation of powerful learned image priors without explicitly defining regularization functions. The overall algorithm alternates between a denoising step that enforces prior constraints and a network optimization step that enforces data consistency, along with a dual variable update to ensure convergence. Through iterative updates, the method progressively refines the reconstruction toward a solution that satisfies both physical consistency and prior knowledge. Results and Discussions The effectiveness of the proposed method is validated through both numerical simulations and physically realistic ultrafast dynamic experiments. In simulation studies, synthetic dynamic scenes consisting of time-varying amplitude and phase distributions are generated, and the corresponding undersampled measurements are obtained using a forward model that closely mimics the actual imaging system. The proposed algorithm successfully reconstructs high-quality complex-amplitude sequences from these measurements, even under high compression ratios and additive noise. Quantitative evaluations using peak signal-to-noise ratio and structural similarity index demonstrate that the proposed method significantly outperforms conventional approaches, including traditional phase retrieval algorithms, deep image prior methods, and plug-and-play methods with simple regularization. In particular, the integration of deep denoising networks enhances noise suppression and detail preservation, leading to improved reconstruction accuracy and structural fidelity. The results also show that different denoisers can be flexibly incorporated into the framework, with advanced models providing superior performance. Robustness analysis further confirms that the method maintains stable reconstruction quality across a wide range of noise levels and undersampling ratios. Even under severe degradation conditions, the reconstructed amplitude and phase images retain clear structural information, and the performance metrics exhibit only gradual degradation without abrupt failure. This indicates that the hybrid prior effectively constrains the solution space and mitigates the ill-posedness of the inverse problem. The convergence behavior of the algorithm is also investigated. The loss function decreases rapidly in the early iterations and stabilizes afterward, indicating efficient convergence. The computational cost per iteration remains moderate, making the method practical for real applications. To further evaluate the applicability of the method in realistic scenarios, it is tested on ultrafast dynamic data generated from a pump-probe imaging system. These data incorporate real physical parameters, system imperfections, and complex temporal evolution, providing a challenging benchmark for reconstruction algorithms. The proposed method accurately recovers both amplitude and phase distributions over time,capturing key features such as the evolution of ablation-induced structures. The reconstructed phase information is further used to infer physical quantities, such as surface deformation, demonstrating the capability of the method to support quantitative analysis of ultrafast phenomena. Overall, the results highlight the advantages of combining physics-based modeling, deep image priors, and plug-and-play denoising. The hybrid framework effectively balances data fidelity and prior constraints,leading to improved reconstruction quality, robustness, and generalization ability compared with existing methods. Conclusions This work presents a hybrid deep prior-based plug-and-play reconstruction algorithm for snapshot broadband coherent diffraction ultrafast imaging. By integrating a physics-driven deep image prior with pretrained deep denoising networks within an alternating optimization framework, the proposed method enables high-fidelity reconstruction of high-dimensional complex-amplitude data from severely undersampled measurements without relying on large-scale training datasets. Extensive experiments demonstrate that the method achieves superior performance in terms of reconstruction accuracy, noise robustness, and structural preservation. It exhibits stable convergence behavior and maintains effectiveness under challenging conditions, including high noise levels and low sampling rates. Moreover, its successful application to physically realistic ultrafast dynamic data confirms its practical utility and potential for real-world imaging systems. The proposed approach provides a flexible and general framework for solving ill-posed inverse problems in ultrafast optical imaging. By bridging physical modeling and data-driven priors, it opens new possibilities for high-resolution, high-speed complex-amplitude imaging and quantitative analysis of dynamic physical processes.