Depth Separable-CNN for Improved Spectral Super-Resolution

Abstract

Hyperspectral Image reconstruction from RGB images is a low-cost and convenient alternative to acquiring hyperspectral images directly. The challenge in estimating the spectral response function and using it for generating the hyperspectral image data is addressed effectively by the use of Deep convolutional networks for the task. Deep networks, however, involve expensive convolutions. To address this problem, we have adapted deep convolution networks into Deep Separable convolution networks. We propose an isotropic network which consists of a cascade of blocks implementing depth separable convolution constructed by a series combination of a spatial convolution followed by convolution along the spectral domain. The two sub-blocks are termed as intra-channel sub-block and intra-pixel sub-block respectively. Intra-channel sub-block helps in capturing spatial correlation and intra-pixel sub-block captures spectral correlations. The explicit use of spectral-domain convolution results in a more direct approach to interpolating information across bands and hence helps in obtaining more accurate Hyperspectral data from RGB images. In this study we have conducted experiments on three benchmark datasets to show that deep separable-CNN achieves better accuracy. These result and the ablation study demonstrate the superiority of our technique.

Keywords

• Convolutional networks
• correlations
• separable networks
• spectral superresolution