Ma Runchuan, He Sailing
National Engineering Research Center for Optical Instruments, Centre for Optical and Electromagnetic Research, College of Optical Science and Engineering, Zhejiang University, Hangzhou, 310058, China.
Department of Electromagnetic Engineering, School of Electrical Engineering, KTH Royal Institute of Technology, Stockholm, SE-100 44, Sweden.
Sci Rep. 2025 May 9;15(1):16253. doi: 10.1038/s41598-025-00877-8.
Hyperspectral imaging and Neural Radiance Field (NeRF) can be combined in powerful ways. With limited hyperspectral images, NeRF can generate images of objects with spectral information from arbitrary viewpoints, which can effectively mitigate defects such as long acquisition time and difficulty in obtaining hyperspectral images. This paper addresses challenges in the application of NeRF methods in the hyperspectral domain, including local errors in convergence caused by noise. Leveraging the characteristics of hyperspectral data, we propose a neural radiance field method employing a multi-channel volume density distribution function. This approach alleviates issues during the generation of neural radiance fields from hyperspectral data, enhancing the robustness of hyperspectral neural radiance field methods across various scenarios, which can help downstream tasks such as discriminating objects more effectively than RGB methods. Experiments demonstrate that the proposed method generates superior hyperspectral images under diverse conditions, with a maximum PSNR 37.66 and a maximum SSIM 0.982.
高光谱成像和神经辐射场（NeRF）可以以强大的方式结合。利用有限的高光谱图像，NeRF可以从任意视角生成具有光谱信息的物体图像，这可以有效缓解诸如采集时间长和获取高光谱图像困难等缺陷。本文探讨了NeRF方法在高光谱领域应用中的挑战，包括由噪声引起的收敛局部误差。利用高光谱数据的特性，我们提出了一种采用多通道体密度分布函数的神经辐射场方法。这种方法缓解了从高光谱数据生成神经辐射场过程中的问题，增强了高光谱神经辐射场方法在各种场景下的鲁棒性，这有助于下游任务，例如比RGB方法更有效地辨别物体。实验表明，所提出的方法在不同条件下生成了 superior 高光谱图像，最大PSNR为37.66，最大SSIM为0.982。 （注：“superior”这里直接保留英文未翻译，因为不太明确其在该语境下准确对应的中文词汇，可能是“更好的”之类意思，需结合具体专业领域进一步确定准确译法）
