Multi-channel volume density neural radiance field for hyperspectral imaging
Runchuan Ma1, Sailing He2,3
1National Engineering Research Center for Optical Instruments, Centre for Optical and Electromagnetic Research, College of Optical Science and Engineering, Zhejiang University, Hangzhou, 310058, China.
Scientific Reports
|May 9, 2025
Summary
This study introduces a novel Neural Radiance Field (NeRF) method for hyperspectral imaging, improving image generation quality and robustness. The enhanced NeRF effectively addresses noise and data limitations, leading to superior hyperspectral image reconstruction.
Area of Science:
- Computer Vision
- Spectroscopy
- Machine Learning
Background:
- Hyperspectral imaging (HSI) captures detailed spectral information but faces challenges like long acquisition times and data scarcity.
- Neural Radiance Fields (NeRF) offer novel view synthesis but require adaptation for hyperspectral data, especially concerning noise and convergence issues.
Purpose of the Study:
- To develop a robust Neural Radiance Field (NeRF) method tailored for hyperspectral imaging.
- To enhance the generation of hyperspectral images from limited data by mitigating noise and improving convergence.
- To improve object discrimination capabilities compared to traditional RGB methods.
Main Methods:
- Proposed a novel NeRF approach utilizing a multi-channel volume density distribution function.
- Leveraged hyperspectral data characteristics to address local convergence errors caused by noise.
- Developed a method to generate neural radiance fields from limited hyperspectral images.
Main Results:
- The proposed method demonstrated superior performance in generating hyperspectral images across diverse conditions.
- Achieved a maximum Peak Signal-to-Noise Ratio (PSNR) of 37.66 and a maximum Structural Similarity Index Measure (SSIM) of 0.982.
- Enhanced robustness of hyperspectral NeRF methods in various scenarios.
Conclusions:
- The developed multi-channel volume density NeRF method effectively overcomes limitations in hyperspectral data acquisition and processing.
- The approach significantly improves the quality and robustness of generated hyperspectral images.
- This advancement holds promise for improving downstream tasks like object discrimination.
