Incorporating Physical Constraints for Color Correction and Spectral Recovery

Kucuk, Abdullah (2025) Incorporating Physical Constraints for Color Correction and Spectral Recovery. Doctoral thesis, University of East Anglia.

Preview

PDF Download (16MB) | Preview

Abstract

RGB cameras sense the colour signal with three device-dependent measurements that differ from the standard Human Visual System (HVS). While standard human colour vision is mediated by cone cells, each RGB camera uses its own device-specific spectral sensitivity functions. Therefore, a colour correction step is necessary to convert these values into standard colour spaces. Sampling light with only three values also loses spectral information about the scene, which is important in many scientific fields. To address this, spectral reconstruction algorithms have been developed to estimate spectra solely from RGB data. However, hyperspectral or multispectral cameras, which capture more comprehensive spectral information, are preferable. Unfortunately, these cameras often encounter resolution limitations due to the constraints of sensing hardware. Traditional pan-sharpening techniques are designed to enhance the spatial resolution of these images by leveraging their higher-resolution panchromatic counterparts. RGB-guided approaches, however, fuse high-resolution RGB images instead of panchromatic images, offering the promise of improved spectral estimation.

This thesis explores the topics of colour correction and spectral recovery algorithms for RGB, multispectral, and hyperspectral cameras by incorporating physical constraints. The contributions are presented in four main chapters as follows.

We first examine colour correction by reviewing existing algorithms for RGB cameras and comparing traditional regression methods with neural network (NN) solutions. Our findings indicate that while NNs outperform simple linear regression, polynomial and root-polynomial regressions still surpass current NN approaches. We then address the issue of exposure invariance in colour correction algorithms and propose an exposure-invariant neural network solution.

Next, we enhance spectral recovery algorithms by incorporating physical constraints. Revisiting the Matrix-R algorithm, we illustrate that any spectrum can be decomposed into its fundamental metamer and metameric black components. The fundamental metamer, which the camera detects, can be directly calculated from RGB values given the camera’s sensitivity functions. The metameric black component, orthogonal to the camera sensor space, does not affect the RGB values. Despite the ability to calculate the correct fundamental metamer for a given RGB, existing spectral reconstruction and pan-sharpening algorithms still exhibit errors in the fundamental metamer part of the signal. The Matrix-R post-processing framework ensures that any spectral reconstruction algorithm has the correct fundamental metamer. Here, we further develop the framework for pan-sharpening methods and include the generalisation for multispectral images. Additionally, we extend the Matrix-R approach and demonstrate that representing spectra in lower dimensions further enhances algorithm performance.

Furthermore, we present another novel pan-sharpening post-correction least-squares solution by upsampling hyperspectral pixels and merging them with high-resolution RGB data. We prove that these post-corrected spectra have the built-in Matrix-R feature and consistently have the correct fundamental metamer. The standalone version of this algorithm also achieves competitive performance compared to the latest state-of-the-art deep learning pan-sharpening solutions.

Finally, we investigate a dual-camera solution for colour correction where high-resolution RGB and corresponding low-resolution multispectral pixels are available, similar to the setup for the pan-sharpening problem. We propose a novel colour correction method using linear regression by integrating upsampled multispectral and the corresponding RGB pixels. In experiments, the technique delivers superior performance than the classical RGB-only colour correction algorithms.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Science > School of Computing Sciences
Depositing User:	Chris White
Date Deposited:	08 Jun 2026 12:05
Last Modified:	08 Jun 2026 12:05
URI:	https://ueaeprints.uea.ac.uk/id/eprint/103312
DOI:

Downloads

Actions (login required)

View Item