[공지사항]을 빙자한 안부와 근황 
Show more

Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Learning Spectral Diffusion Prior for Hyperspectral Image Reconstruction

Created by
  • Haebom

Author

Mingyang Yu, Zhijian Wu, Dingjiang Huang

Outline

This paper addresses the problem of hyperspectral image reconstruction, which reconstructs 3D hyperspectral images (HSIs) from disturbed 2D measurements. Pointing out that existing deep learning-based methods have difficulty in accurately capturing high-frequency details, we propose a spectral spread dictionary (SDP) implicitly learned from hyperspectral images using a diffusion model. By leveraging the detail restoration ability of the diffusion model, the learned dictionary is applied to the HSI model to significantly improve its performance. In addition, to further enhance the effectiveness of the learned dictionary, a spectral dictionary injection module (SPIM) is proposed to induce the model to dynamically restore HSI details. The proposed method is evaluated on two representative HSI methods, MST and BISRNet, and the performance is improved by about 0.5 dB over the conventional networks.

Takeaways, Limitations

•
Takeaways:
◦
We show that hyperspectral image reconstruction performance can be improved by learning a spread spectral dictionary (SDP) using a diffusion model.
◦
We propose that the effectiveness of learned dictionaries can be further enhanced by using the Spectral Pre-Injection Module (SPIM).
◦
Contributes to the field of hyperspectral image reconstruction by achieving approximately 0.5 dB improved performance compared to existing methods.
•
Limitations:
◦
The performance improvement of the proposed method is relatively small, at 0.5 dB.
◦
Experiments on diverse hyperspectral image datasets may be lacking.
◦
A detailed description of the design and operation of the SPIM module may be lacking.
◦
There may be a lack of analysis of the computational cost and time required to learn diffusion models.
👍
Made with Slashpage