Auto-Encoder Guided Attention Based Network for Hyperspectral Recovery from Real RGB Images

Hyperspectral reconstruction from RGB images have been proposed as an alternative approach to overcome the limited scope of conventional hyperspectral imaging. With the advancement in convolutional neural networks (CNNs), this approach has recently gained attention. However, most of the existing state-of-the-art frameworks focuses on hyperspectral reconstruction from clean RGB images i.e. with no noise and degradation. This limits the applicability of the current deep learning frameworks on real-world RGB images. Thus, in this paper we propose a novel deep learning framework for robust real RGB (with noise and degradation) for hyperspectral reconstruction. The proposed framework is motivated towards the extraction of noise-free features from real RGB images crucial for hyperspectral reconstruction. This is achieved with the use of a deep convolutional auto-encoder (DCAE) module and subsequent utilization of these noise-free features in an attention-based deep spectral back-projection network (DSBPN) for hyperspectral reconstruction. The proposed framework (DCAE-DSBPN) is trained in an end-to-end manner with joint optimization of denoising loss and hyperspectral reconstruction loss. Experimental results demonstrates that the proposed framework outperforms the existing state-of-the-art methods.