Study of a deep learning-based method for improving the spectral resolution of the spectral scanning hyperspectral imaging system via synthetic spectral image data

Abstract

Hyperspectral imaging (HSI) techniques, measuring spatial and spectral information, have shown the ability to identify targets based on their spectral features. Among many HSI methods, a spectral scanning HSI method implemented using a tunable filter has been widely used in various applications due to wide-area HSI capability and cost-effectiveness. However, the limitation of the spectral scanning method is poor spectral resolution compared to other spectral imaging methods using dispersive materials. To overcome this limitation, we exploited a recently developed deep-learning model that retrieves multispectral information from an red, green, and blue image. Moreover, this study proposed that a color chart consisting of 18 colors could be a standard target for training the deep-learning model under various spectral scanning HSI conditions. The simulation work was performed to demonstrate the feasibility of the proposed method using synthetic hyperspectral images. Realistic synthetic data was prepared using spectral data obtained via a spectrometer (ground-truth data) and artificial filters emulating a liquid-crystal tunable filter. We found that the deep-learning model trained via a supervised learning approach using synthetic hyperspectral images successfully retrieved high-resolution spectral data. In addition, the trained deep-learning model retrieved robust spectral profiles of random colors which were not used in the training process. Collectively, the deep learning-based spectral scanning method could improve the spectral resolution of the imaging system, and the color chart would be a good and practical standard training target for the deep learning model.