| 研究実績の概要 |
This research proposal aims to optimize the filter response for multispectral-to-hyperspectral reconstruction using deep neural networks. In FY2019, towards our ultimate goal, we have first collected a spectral database with more than 200 scenes, under 5 types of illuminants (Halogen lamp, Xenon lamp, LED lamp, Fluorescent lamp and Incandescent lamp). This database is being expanded further to include outdoor scenes under natural illumination. In the aspect of customized DNN for hyperspectral reconstruction, we have developed a multi-level and multi-scale spatial and spectral fusion CNN, which is shown to be effective.
We have designed the optimal filter response for hybrid-fusion based spectral reconstruction, and the advantages of the designed filters are proved.
|
| 今後の研究の推進方策 |
In FY2020, the major research tasks include (i). To further enrich the spectral database by capturing outdoor scenes under natural daylight and skylight illumination. This database will be released for academic usage. (ii). To systematically study the task of optimal filter response design in the mosaiced and non-mosaiced case, and examine the effect of illumination metamerism. A paper introducing these research findings will be submitted to IEEE TPAMI. (iii). Rather than designing a filter array from scratch, it is planned to further enhance the existing Bayer pattern for spectral reconstruction, by designing a single filter that will act on the three color channels simultaneously. A paper on this topic will be submitted to CVPR2021.
|
