Invertible Diffusion Models for Compressed Sensing.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Deep neural networks (NNs) enhance image compressed sensing (CS) but require training from scratch, limiting deployment.
• Current diffusion model approaches for CS reconstruction face challenges with slow inference and limited adaptability.

Purpose of the Study:

• To propose Invertible Diffusion Models (IDM), an efficient, end-to-end diffusion-based method for image CS reconstruction.
• To overcome the limitations of existing CS reconstruction methods, including slow inference and the need for scratch training.

Main Methods:

• IDM repurposes a diffusion sampling process as a reconstruction model, fine-tuned end-to-end for direct image recovery from CS measurements.
• A novel two-level invertible design transforms the multi-step sampling process and noise estimation U-Net into invertible networks, reducing GPU memory usage by up to 93.8%.
• Lightweight modules are developed to integrate CS measurements into the noise estimator, enhancing reconstruction.

Main Results:

• IDM demonstrates superior performance over state-of-the-art CS networks, achieving up to 2.64 dB higher Peak Signal-to-Noise Ratio (PSNR).
• Compared to DDNM, IDM provides up to 10.09 dB PSNR gain and is 14.54 times faster.
• Significant reduction in GPU memory requirements during training due to the invertible design.

Conclusions:

• IDM presents a novel and efficient approach to image CS reconstruction using diffusion models.
• The proposed method achieves state-of-the-art reconstruction quality and significantly accelerates inference speed.
• IDM's memory efficiency and adaptability make it a promising solution for practical CS applications.