Evaluation of the Effect of Light Source Spectrum on the Estimation Accuracy of High Dimensional Spectral Images Using Neural Networks Based on RGB Images

Recently, hyperspectral images are becoming widespread and being applied to several fields. However, the expensive cost of a hyperspectral camera poses a problem for introducing it to other fields. Contrastingly, RGB images can be easily obtained using digital cameras, including smartphones. Hence, we have proposed an estimation method of hyperspectral images from RGB images using a neural network. In our proposed method, spectral characteristics of light sources strongly affect the estimation accuracy of hyperspectral images. In this study, we evaluate the effect of light source spectrum on the estimation accuracy in both outside and inside situations. Two models are constructed using neural networks independently by input images shot in an outside, and inside situation. Two images of a cushion taken outside and inside are respectively input into the constructed models, and the hyperspectral image is estimated for each. Comparing estimated images in outside and inside situations, estimated images in an outside situation showed high quality, accuracy, and high similarities. Considerable noise was included in images shot in an inside situation, and the estimation accuracy was found to be lower than that of the corresponding images shot in outside situations. Thus, light source spectrum has a huge effect on the estimation accuracy of hyperspectral images, and it is best for estimating hyperspectral images to input images shot under solar light using neural networks.