2020 Fiscal Year Annual Research Report
Data-driven Filter Design and Implementation for Snapshot Hyperspectral Imaging
| Project/Area Number |
19K20307
|
|---|
| Research Institution | National Institute of Informatics |
|---|
Principal Investigator |
鄭 銀強 国立情報学研究所, コンテンツ科学研究系, 准教授 (30756896)
|
|---|
| Project Period (FY) |
2019-04-01 – 2021-03-31
|
| Keywords | spectral imaging / filter design / deep learning |
|---|
| Outline of Annual Research Achievements |
In FY2020, two researches have been conducted. 1. Optimal filter selection from an existing camera spectral response dataset. Through exhausitive evaluaiton, it has been shown that spectral response curves indeed pose obvious effect on spectral reconstruction. Also, through deep learning, an algorithm to identify the best filters has been proposed. 2. IR-cut filter design for optimal spectral reconstruction from RGB observations. Rather than using existing filters, this research is to design new filters for further improved spectral reconstruction. Compared with existing three-channel design method, the proposed IR-cut filter design method only needs to customize one filter, and the production cost can be much reduced, without sacrificing much spectral reconstruction accuracy.
Through the researches in FY2019 and FY2020, systematic researches on reconstruction-based spectral imaging have been conducted, including more efficient network framework for RGB-to-Spectrum umsampling, systematic evaluation on the effect of filters in spectral reconstruction, and the optimal selection of existing filters for improved spectral reconstruciton, as well as the optimal design of filters for spectral reconstruction with the highest accuracy. Academic publications regarding these research achievements have been accepted to top international conferences and journals. It is planned to utilize these research results to assist spectral camera development in the near future.
|
