Reconstruction of hyperspectral images from RGB images

Abstract

This paper accounts for the problem of construction of hyper-spectral (hs) images from RGB-images, i.e recovery of the whole spectral details/signature from a three-channel RGB image. The dataset used in this paper consists of ‘clean’ images, that are images without noise. There are 450 clean images along with correlative 450 hyperspectral images and all the pages are in PNG (.png) format. We approached this problem using 3 models Convhs_5, Enhanced-ResNet and Dense-HSCNN (D-HSCNN). These models increase in complexity from Convhs_5 to Dense-HSCNN. In the evaluation phase, Convhs_5 achieved average results with not so much clarity, whereas the best model came out to be Enhanced-ResNet with satisfying results, however, due to insufficiency of computational power/resources, we were not able to get the expected results from the Dense-HSCNN model. Whilst having more number of images in the dataset and more computational power can reduce the loss in all of the models furthermore.

Advanced Disease Monitoring and Severity Quantification for Greenhouse Management Using Multispectral Imaging

“…Nguyen et al [6] explored hyperspectral imaging for early plant viral disease detection, facing challenges with a constrained sample size. Pushparaj et al [7] makes a substantial contribution by conducting a comparative analysis of several Convolutional Neural Network (CNN)-based methods employed in hyperspectral image reconstruction. The evaluation utilizes the NTIRE 2020 challenge dataset and specifically examines the performance of distinct models, encompassing a 5-layer basic CNN, Enhanced-ResNet with 10 layers, and Dense-HSCNN.…”

“…The research is positioned as a valuable resource for practitioners seeking optimal strategies for hyperspectral image reconstruction. Pushparaj et al [7] conduct a comparative analysis of CNN-based methods in hyperspectral image reconstruction, contributing valuable insights. L. Yan et al [8] presents a pioneering CNN based framework designed for the recovery of hyperspectral information from RGB images.…”