Deep Unfolding Architecture Based on Generative Prior Diffusion for Image Compressive Sensing

Deep neural networks have led to the development of various data-driven architectures for image compressive sensing (CS), generally categorized into closed-box models and deep unfolding networks (DUNs). DUN approaches have garnered attention due to their interpretability. However, most existing methods emphasize the design of iterative networks while insufficiently addressing the use of image priors. This study introduces a novel deep unfolding architecture for image CS, called GPD-CS, which integrates generative priors (GPs) and diffusion models (DMs). The proposed method involves a two-stage training process. In the first stage, a modified variational autoencoder is designed to capture compact prior representations of natural images, and a dynamic proximal mapping module is incorporated within the DUN to integrate prior information. During the second stage, the GP learning model’s parameters are fixed, and a lightweight DM is employed to infer compact representations, which are then embedded into the iterative reconstruction process. Experiments validate the superior reconstruction performance of the proposed approach.