Introduction
Image Compressive Sensing (CS) is a critical technique for reconstructing high-resolution images from a fraction of the original data. This ill-posed inverse problem is traditionally approached via two methods: Deep Unfolding Networks (DUNs), which are efficient but limited by shallow priors, and Diffusion Models, which offer powerful generative priors but suffer from high computational costs and slow inference.
The proposed DMP-DUN framework bridges this gap by embedding a pre-trained diffusion model into a novel iterative algorithm, then unfolding it into a neural network. This approach achieves state-of-the-art reconstruction quality in as few as 2 to 10 steps, drastically reducing the complexity of traditional diffusion-based reconstruction.
Methodology
The authors introduce a two-stage solution to integrate diffusion priors into the DUN framework.
Diffusion Message Passing (DMP)
The core of the method is the DMP algorithm, inspired by Approximate Message Passing (AMP). DMP embeds a single step of a reverse diffusion process into each iteration, ensuring that intermediate reconstructions remain on the manifold of natural images.
DMP-DUN Architecture
The DMP algorithm is unfolded into a deep neural network with several key innovations:
- Onsager Term Approximation: The computationally expensive divergence calculation in AMP is replaced with a lightweight ResBlock, accelerating the network.
- Learnable Hyperparameters: Time steps and scaling parameters are learned end-to-end rather than being fixed.
- DMP-DUN+: An enhanced version where the pre-trained diffusion model's weights are fine-tuned during the CS training process.
- Head/Tail ResBlocks: These modules map raw measurements into the diffusion model's latent space and convert the final output back to image format.
Experimental Results
The model was trained on the COCO2017 dataset and evaluated on standard benchmarks like Set11, Urban100, and BSDS500.
Performance and Efficiency
DMP-DUN+ demonstrates a superior trade-off between reconstruction quality and computational cost. As shown in Figure 1, the 4-step DMP-DUN+ achieves higher PSNR than competing methods while maintaining significantly lower FLOPs compared to traditional diffusion-based methods like DDNM.
Visual Fidelity
Qualitative analysis confirms that DMP-DUN+ recovers sharp, realistic textures that other DUNs often blur. On the Urban100 dataset at a 10% CS ratio, DMP-DUN+ achieved a PSNR of 31.39 dB, outperforming the previous best (PRL-RND+) by over 2.6 dB.
Conclusion
DMP-DUN effectively leverages the "brainpower" of diffusion models within a high-speed unfolding framework. While training the unfrozen DMP-DUN+ is computationally intensive, the resulting inference speed and reconstruction quality represent a significant advancement in image compressive sensing. Future work may extend this framework to specialized domains such as medical imaging or snapshot compressive imaging.
Summary updated