An efficient CNN for spectral reconstruction from RGB images

TL;DR

This paper introduces a moderately deep CNN model for spectral reconstruction from RGB images, significantly improving performance on standard benchmarks compared to previous methods.

Contribution

It presents a new CNN architecture that outperforms existing shallow and deep models in spectral super-resolution tasks.

Findings

Achieved superior performance on ICVL, CAVE, and NUS benchmarks.

Outperformed previous shallow and deep learning methods.

Demonstrated the effectiveness of a moderately deep CNN architecture.

Abstract

Recently, the example-based single image spectral reconstruction from RGB images task, aka, spectral super-resolution was approached by means of deep learning by Galliani et al. The proposed very deep convolutional neural network (CNN) achieved superior performance on recent large benchmarks. However, Aeschbacher et al showed that comparable performance can be achieved by shallow learning method based on A+, a method introduced for image super-resolution by Timofte et al. In this paper, we propose a moderately deep CNN model and substantially improve the reported performance on three spectral reconstruction standard benchmarks: ICVL, CAVE, and NUS.