Spectral Compressive Imaging via Unmixing-driven Subspace Diffusion Refinement

Authors: Haijin Zeng, Benteng Sun, Yongyong Chen, Jingyong Su, Yong Xu

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on the standard KAIST and zero-shot datasets NTIRE, ICVL, and Harvard show that PSR-SCI enhances overall visual quality and delivers PSNR and SSIM results competitive with state-of-the-art diffusion, transformer, and deep-unfolding baselines. This framework provides a robust alternative to traditional deterministic SCI reconstruction methods.
Researcher Affiliation Academia 1 Harvard University, 2 Harbin Institute of Technology (Shenzhen)
Pseudocode Yes Algorithm 1 Predict-and-subspace-refine diffusion sampling.
Open Source Code Yes Code and models are available at https://github.com/SMARK2022/PSR-SCI.
Open Datasets Yes Experimental results on the standard KAIST and zero-shot datasets NTIRE, ICVL, and Harvard show that PSR-SCI enhances overall visual quality and delivers PSNR and SSIM results competitive with state-of-the-art diffusion, transformer, and deep-unfolding baselines. CAVE dataset Park et al. (2007). KAIST Choi et al. (2017).
Dataset Splits Yes Similar to most existing methods Meng et al. (2020a); Hu et al. (2022); Huang et al. (2021); Cai et al. (2022d), we select 10 scenes with a spatial size of 256 256 and 28 bands from KAIST Choi et al. (2017) as the simulation dataset for testing. Meanwhile, we also select 5 MSIs with a spatial size of 660 660 and 28 bands, captured by the CASSI system as the real dataset Meng et al. (2020a), and then crop the MSIs into data blocks of size 256 256 for testing.To evaluate generalization performance of our approach, we test it on several zero-shot MSI datasets, including ICVL, NTIRE, and Harvard, which were not used during training.
Hardware Specification Yes All experiments are conducted with data paralleling on a server equipped with 4 RTX 3090 GPUs, using Python 3.9.19, PyTorch 2.2.0+cu121, and CUDA 12.2.
Software Dependencies Yes All experiments are conducted with data paralleling on a server equipped with 4 RTX 3090 GPUs, using Python 3.9.19, PyTorch 2.2.0+cu121, and CUDA 12.2.
Experiment Setup Yes For the URSe training, we initialize the model randomly. Training is performed using the Adam optimizer with a learning rate (lr) of 0.02 and a batchsize of 8 for 200 epochs. During the first 30% of the epochs, an additional visual enhancement regularization term based on the mean squared error (MSE) between encoded images and pseudo-RGB images is included. The first 10 epochs employ a linear learning rate warmup strategy, and over the next 100 epochs, the lr is gradually reduced to 0.002 to achieve higher performance. The initial training of URSe takes approximately 2.5 hours.