Spectral Estimation from Actual Color Images based on Deep Neural Network

Abstract

Spectral information can be recovered from color images to preserve high spectral and spatial resolutions simultaneously. Synthetic color images are generally used in the present methods. However, the synthetic color images might not conform to the data distribution of actual color images. In this study, a real dataset is constructed with actual color images acquired with color cameras. Besides, a deep neural network is designed using residual neural network, to retrieve the spectral reflectances from the RGB values of the color images. The proposed network is verified via a 5-fold cross validation, and the findings indicate that the proposed network can obtain higher accuracy than the two existing spectral estimation algorithms with respect to the constructed dataset.