NTIRE 2022 Spectral Recovery Challenge and Data Set

Abstract

This paper reviews the third biennial challenge on spectral reconstruction from RGB images, i.e., the recovery of whole-scene hyperspectral (HS) information from a 3-channel RGB image. This challenge presents the "ARAD 1K" data set: a new, larger-than-ever natural hyperspectral image data set containing 1,000 images. Challenge participants were required to recover hyperspectral information from synthetically generated JPEGcompressed RGB images simulating capture by a known calibrated camera, operating under partially known parameters, in a setting which includes acquisition noise. The challenge was attended by 241 teams, with 60 teams competing in the final testing phase, 12 of which provided detailed descriptions of their methodology which are included in this report. The performance of these submissions is reviewed and provided here as a gauge for the current stateof-the-art in spectral reconstruction from natural RGB images.

Fast-n-Squeeze: towards real-time spectral reconstruction from RGB images

“…As is possible to see, the proposed method with 198.45 FPS is considerably faster than the other methods. This number is higher than the 104.71 FPS reported in [5] because we optimize the code efficiency by asynchronously running the forward pass for the two SqueezeNet models. Compared to deep learningbased methods that run in GPUs, it is an order of magnitude faster than HRNet and three orders of magnitude faster than AWAN.…”

“…We train and evaluate our Fast-n-Squeeze on the Arad 1K Hyperspectral Database provided by NTIRE 2022 challenge [5].…”

“…This paper describes our solution to the NTIRE 2022 Spectral Reconstruction Challenge [5], in which each participant is provided with a set of RGB-spectral image pairs for development, training, and tuning, and the final assessment is conducted on a blind test set. As we will discuss later on, existing methods for spectral reconstruction are traditionally based on defining sparse-representation dictionaries to encode the transformation from RGB to spectral domain.…”

Maize disease detection based on spectral recovery from RGB images

“…Owing to our goal is to recovery HSIs from natural RGB images and the wavelength of natural RGB images ranges from about 400 -700 nm. For the purpose of reducing training cost and improving training efficiency, the images were resampled to 31 spectral bands in the visual range from 400 nm to 700 nm with a spectral resolution of 10 nm (Arad et al (2022)). In this study, the images of maize were captured at a distance of 1-1.5 m. A neutral reference panel with 99% reflection efficiency was used to perform spectral calibration.…”

NTIRE 2022 Challenge on Night Photography Rendering

“…This challenge is one of the NTIRE 2022 associated challenges: spectral recovery [6], spectral demosaicing [5], perceptual image quality assessment [26], inpainting [46], efficient super-resolution [35], learning the super-resolution space [41], super-resolution and quality enhancement of compressed video [55], high dynamic range [44], stereo super-resolution [52], burst super-resolution [8].…”