Deep Learning-Based Exposure Asymmetry Multispectral Reconstruction from Digital RGB Images

Multispectral reconstruction is an important way to acquire spectral images with a high spatial resolution as snapshots. Current deep learning-based multispectral reconstruction models perform well under symmetric conditions, where the exposure of training and testing images is consistent. However, further research has shown that these models are sensitive to exposure changes. When the exposure symmetry is not maintained and testing images are input into the multispectral reconstruction model under different exposure conditions, the reconstructed multispectral images tend to deviate from the real ground truth to varying degrees. This limitation restricts the robustness and applicability of the model in practical scenarios. To address this challenge, we propose an exposure estimation multispectral reconstruction model of EFMST++ with data augmentation and optimized deep learning architecture, where Retinex decomposition and a wavelet transform are introduced into the proposed model. Based on the currently available dataset in this field, a comprehensive comparison is made between the proposed and existing models. The results show that after the current multispectral reconstruction models are retrained using the augmented datasets, the average MRAE and RMSE of the current most advanced model of MST++ are reduced from 0.570 and 0.064 to 0.236 and 0.040, respectively. The proposed method further reduces the average MRAE and RMSE to 0.229 and 0.037, with the average PSNR increasing from 27.94 to 31.43. The proposed model supports the use of multispectral reconstruction in open environments.