Flow-Matching Guided Deep Unfolding for Hyperspectral Image Reconstruction

TL;DR

The Flow-Matching-guided Unfolding network is proposed, which is the first to integrate flow matching into HSI reconstruction by embedding its generative prior within a deep unfolding framework and introduces a mean velocity loss that enforces global consistency of the flow, leading to more robust and accurate reconstruction.

Abstract

Hyperspectral imaging (HSI) provides rich spatial-spectral information but remains costly to acquire due to hardware limitations and the difficulty of reconstructing three-dimensional data from compressed measurements. Although compressive sensing systems such as CASSI improve efficiency, accurate reconstruction is still challenged by severe degradation and loss of fine spectral details. We propose the Flow-Matching-guided Unfolding network (FMU), which, to our knowledge, is the first to integrate flow matching into HSI reconstruction by embedding its generative prior within a deep unfolding framework. To further strengthen the learned dynamics, we introduce a mean velocity loss that enforces global consistency of the flow, leading to a more robust and accurate reconstruction. This hybrid design leverages the interpretability of optimization-based methods and the generative capacity of flow matching. Extensive experiments on both simulated and real datasets show that FMU significantly outperforms existing approaches in reconstruction quality. Code and models will be available at https://github.com/YiAi03/FMU.

Figures (6)

• Figure 1: PSNR-FLOPS comparison with recent HSI reconstruction methods. Input size is 256$\times$256 for FLOPs computation.
• Figure 2: System Overview: Classical CASSI vs. Optical Filters-Based HSI Systems
• Figure 3: Overall pipeline of our method. The measurement $\bm y$ passes through FMU with N stages and finally get the output reconstruction. In each stage, there is a GAP projection and a U-shaped denoiser consists of Trident Transformers, which is assisted with the prior from flow matching.
• Figure 4: Two-phase training procedure of our method. In first phase, we train a latent encoder to learn knowledge from clean HSIs; and in second phase we fix the latent encoder and train flow matching with the prior from clean HSI to generate prior conditioned on measurement.
• Figure 5: Qualitative comparison on simulated data. From top to bottom, each row visualizes the reconstructed channels at wavelengths of 481.5 nm, 522.5 nm, 575.5 nm, and 648.0 nm in scene 2. The top left part shows the measurement, the RGB reference image of the original HSI and the spectral density curves within the yellow region of interest.