PaperlayerView exact revision on arXiv

Source-linked AI summary

A Perceptually Inspired Variational Framework for Color Enhancement

Rodrigo Palma-Amestoy, Edoardo Provenzi, Marcelo Bertalmío, Vicent Caselles

arXiv:2511.23329v1cs.CV

TL;DR

Color-correction models inspired by human vision can be difficult to characterize in terms of image features such as contrast and dispersion. This paper formulates phenomenological requirements for variational energies and studies their minimizers; tests confirm enhancement and color-cast removal, while noting that existing noise can be amplified.

  • Problem

    Color-correction models inspired by vision can be difficult to characterize in terms of contrast and dispersion, while separating illuminant from reflectance is ill-posed without constraints not met by all images.

    Source: §2.1 · p. 4

  • Method

    The authors translate color-vision phenomenology into assumptions for energy functionals and minimize them using iterative gradient descent.

    Source: §9 · pp. 24–25

  • Results

    Tests confirm that the proposed algorithms can enhance details and remove color cast without introducing noise, although already present noise can eventually be amplified.

    Source: §9 · p. 25

  • Takeaways & Limitations

    The energy assumptions expose minimization as balancing local nonlinear contrast enhancement against control of dispersion around middle gray and input values.

    Source: §9 · p. 25

    Image-dependent tuning could possibly improve enhancement, but is beyond the paper’s scope and remains an open problem.

    Source: §9 · p. 25

Abstract

from arXiv · show

Basic phenomenology of human color vision has been widely taken as an inspiration to devise explicit color correction algorithms. The behavior of these models in terms of significative image features (such as contrast and dispersion) can be difficult to characterize. To cope with this, we propose to use a variational formulation of color contrast enhancement that is inspired by the basic phenomenology of color perception. In particular, we devise a set of basic requirements to be fulfilled by an energy to be considered as `perceptually inspired', showing that there is an explicit class of functionals satisfying all of them. We single out three explicit functionals that we consider of basic interest, showing similarities and differences with existing models. The minima of such functionals is computed using a gradient descent approach. We also present a general methodology to reduce the computational cost of the algorithms under analysis from ${\cal O}(N^2)$ to ${\cal O}(N\log N)$, being $N$ the number of input pixels.

1 Introduction

The paper translates basic phenomenological properties of color vision into requirements for variational energy functionals, rather than designing correction algorithms directly from visual-system features. It identifies a class of functionals and analyzes three examples, their behavior, and ways to reduce computational cost.

1 Introduction

  • Neurophysiological processes underlying color vision remain insufficiently understood, so most models describe macroscopic visual characteristics rather than reproduce neural activity.The paper therefore treats perceptual models as descriptive rather than as deterministic accounts of vision.Source: §1 · p. 2
  • The framework translates basic color-perception phenomenology into energy requirements whose minimization produces color-corrected images following some basic HVS behavior.The variational formulation is intended to clarify effects on image features such as contrast and dispersion.Source: §1 · p. 2
  • The authors identify a class of functionals satisfying the perceptual assumptions and the logarithmic response to light stimuli expressed by Weber-Fechner’s law.Their Euler-Lagrange equations support gradient-descent algorithms.Source: §1 · p. 2
  • Three selected energies are theoretically and qualitatively compared with Retinex and ACE, and a general procedure reduces algorithmic cost from O(N^2) to O(N log N).N denotes the number of input pixels.Source: §1 · p. 3
  • The paper proceeds from HVS phenomenology and energy assumptions to example energies, analysis, comparisons, computational reduction, and tests.Proofs are placed in the appendices.Source: §1 · p. 3

2 Basic HVS phenomenological properties

The paper frames color constancy, local contrast enhancement, and visual adaptation as three basic phenomenological properties of human color perception. These properties motivate linking channel contrast, spatial context, and scene-average intensity to color correction.

2 Basic HVS phenomenological properties

  • The paper treats color constancy, local contrast enhancement, and visual adaptation as the three basic phenomenological properties motivating its framework.Source: §2 · p. 3

2.1 Human color constancy

  • Human color constancy is imperfect and context-sensitive; the paper links color casts to uneven channel standard deviations and proposes reducing their differences through contrast enhancement.For the blue-cast example, the reported standard deviations are σR = 17.59, σG = 27.03, and σB = 130.56.Source: §2.1 · pp. 3–4

2.2 Local contrast enhancement

  • Chromatic induction and other local-contrast effects depend on spatial context, with induction strength decreasing monotonically with Euclidean distance between areas.Source: §2.2 · p. 4

2.3 Visual adaptation

  • Visual adaptation narrows the usable intensity range around a scene’s average, motivating the gray-world assumption that a scene’s average color is gray.The HVS can adapt across an intensity range on the order of 10^10, but not simultaneously.Source: §2.3 · p. 5

3 The basic set of assumptions

The assumptions define a discrete-image setting and an energy combining spatially dependent local contrast enhancement with global control of departure from the input and middle gray. Contrast terms are constrained by symmetry, homogeneity, and a generalized Weber-Fechner condition.

3 The basic set of assumptions

  • The framework uses discrete RGB images on a normalized [0,1] range, processes scalar channels independently, and extends images by reflection and periodicity.This extension defines the working domain for the subsequent assumptions.Source: §3 · p. 5

3.1 The first assumption: General structure of the energy functional

  • The energy combines a spatially varying term for local contrast enhancement with global dispersion control around both the input values and middle gray, 1/2.The dispersion term uses separate functions to measure departure from the input and from 1/2.Source: §3.1 · pp. 6–7
  • The contrast energy uses a symmetric inverse-contrast function and a positive symmetric kernel to weight mutual pixel influence.The inverse-contrast function decreases as the smaller intensity falls or the larger intensity rises; the kernel localizes the contrast computation.Source: §3.1 · p. 6

3.3 Contrast functions satisfying the Weber-Fechner’s law

  • The inverse-contrast function must be homogeneous, and the generalized Weber-Fechner condition restricts it to a non-decreasing function of min(a,b)/max(a,b).Degree-zero homogeneity makes the contrast depend on an intensity ratio rather than an overall intensity scaling.Source: §3.2 · pp. 7–8, §3.3 · p. 8

4 Explicit energy functionals complying with our set of assumptions

The section gives explicit energy terms satisfying the paper’s assumptions: three inverse-contrast choices and dispersion terms that control departure from the original image and middle gray. The entropic dispersion term is selected for compatibility with the homogeneity of the contrast variations.

4.1 Three basic inverse contrast terms

  • The contrast terms use three increasing functions of min(I(x), I(y))/max(I(x), I(y)), whose minimization enhances contrast.The choices are identity, logarithm, and minus Michelson contrast.Source: §4.1 · p. 9
  • The logarithmic choice turns intensity ratios into differences in log intensities, while the minus-Michelson choice is expressed through the integral of Michelson contrast.Source: §4.1 · p. 9

4.2 Entropic dispersion term

  • Dispersion terms attach the enhanced image to both the original intensities and the assumed average illumination 1/2; a quadratic distance is the simplest option.Source: §4.2 · p. 10
  • The entropic dispersion term is proposed to remain coherent with the degree -1 homogeneity of the contrast-term variations.Its relative-entropy distances are measured between the output and 1/2, and between the original image and output.Source: §4.2 · p. 10
  • Unlike the quadratic dispersion integrands, the entropic integrands are not symmetric about their minima.Source: §4.2 · pp. 10–11, Figure 2 · p. 10

5 Minimization of the energy functionals Ef

The section establishes minima for the energy functionals and derives their variations for gradient-descent minimization. It regularizes the nondifferentiable contrast terms and uses a positive lower bound on intensities to avoid singularities.

5 Minimization of the energy functionals Ef

  • The minimization is restricted to intensities at least ρ, avoiding singularities from logarithmic contrast or entropic dispersion terms.Source: §5 · p. 11
  • The energy combining a continuous contrast term with entropic dispersion has a minimum in F_ρ; an analogous existence result holds with quadratic dispersion.Source: §5 · p. 11
  • Because the contrast terms are not convex and involve a nondifferentiable absolute value, the method regularizes that basic variable before gradient descent.The stated regularization is convex, differentiable, even, and approximates the absolute value as ε tends to zero.Source: §5 · pp. 11–12
  • The entropic dispersion derivative has degree -1 homogeneity in I(x), matching the variations of the three contrast terms.Source: §5 · p. 13
  • A semi-implicit discrete gradient descent in log I is used to search for minima, with a relative-entropy metric rather than the usual quadratic-distance metric.Source: §5 · p. 14

6 Features of the three computational algorithms and comparison with existing models

The three algorithms share a structure but differ in their response across intensity ranges. The logarithmic algorithm is essentially identical to ACE, while the identity-based algorithms are described as novel and as symmetrized Retinex variants.

6 Features of the three computational algorithms and comparison with existing models

  • The logarithmic algorithm is expected to enhance contrast strongly in dark areas and weakly in bright zones, and the authors report that empirical tests verify this behavior.Source: §6 · p. 15

6.1 Correspondence between Elog w,ϵ and the ACE model

  • The logarithmic-energy algorithm and ACE are essentially identical because the sigmoid in ACE approximates the sign function; ACE uses it to avoid abrupt color corrections.Source: §6.1 · p. 16

6.2 Qualitative similarities and differences between Eid

  • The identity-energy models are described as novel, and their corresponding output images are reported to be very similar.Source: §6.2 · p. 16
  • Unlike ratio-reset Retinex, the identity-energy equations use signed ratios and can either increase or reduce pixel intensity, motivating their description as symmetrized Retinex models.Source: §6.2 · pp. 16–17

7 A general strategy for computational complexity reduction

The algorithms’ direct computation has quadratic complexity in the number of pixels. The proposed approximation separates intensity dependencies into convolutions computed by FFT, reducing complexity to O(N log N) when the approximation degree is much smaller than N.

7 A general strategy for computational complexity reduction

  • The original computation is O(N^2), and processing a high-resolution image can take hours on a standard PC.Source: §7 · p. 17
  • The proposed speedup uses mathematical properties of the r-functions rather than local sampling, preserving chromatic information while avoiding sampling noise.Source: §7 · p. 17
  • Separating the dependencies on I(x) and I(y) with a polynomial approximation expresses the computation as a sum of convolutions, which FFT computes in O(N log N).When n << N, the resulting complexity is O(N log N).Source: §7 · pp. 17–18
  • Polynomial approximation may be unsuitable when the surface of r(I(x), I(y)) has discontinuities or strong gradients; Fourier expansion may then be more suitable but analytically more complicated.Source: §7 · p. 18

8 Tests

Tests show contrast enhancement, color-cast removal, and equalization of underexposed and overexposed images; the energies differ in their treatment of dark and bright regions. Gamma transformation strengthens dark-region enhancement, cognitive-image outputs match reported experimental evidence, and optional noise control works best on very dark images.

8 Tests

  • The tests use a normalized contrast weight and gradient descent, with parameters controlling dispersion, attachment to input data, step size, and signum regularization.The signum function is regularized with an arctangent approximation to avoid unwanted noise amplification.Source: §8 · p. 18

8.1 Some experimental results

  • Tests confirm contrast enhancement, color-cast removal, and equalization of underexposed and overexposed images; E_id and E_-M behave similarly, while E_log favors dark regions and E_id bright areas.The algorithms are evaluated on images with underexposure, a blue cast, and overexposure.Source: Figure 5 · p. 19, Figure 6 · p. 19, §8.1 · p. 20
  • On the overexposed test image, Retinex further increases overexposure, whereas the algorithms for equations (26) and (28) equalize the image.Source: §8.1 · p. 20, Figure 8 · p. 20

8.2 Gamma transformation of the basic contrast variable

  • Replacing the basic contrast-regularized variable with a gamma-transformed one yields energy functionals that comply with all assumptions; gamma-transformed algorithms enhance dark-area contrast more than the originals while maintaining good enhancement in bright zones.Source: §8.2 · pp. 21–22

8.3 Tests on cognitive images

  • Outputs on Mach bands and simultaneous-contrast images are reported as coherent with experimental evidence.Source: §8.3 · p. 22

8.4 A pre and post-processing strategy for noise control

  • For noisy very dark images, extrema-killer filtering before enhancement and restoring removed details afterward works satisfactorily, but may blur brighter or more detailed images.The authors leave this strategy optional for users to select according to image needs or features.Source: §8.4 · p. 24

9 Conclusions and future works

The paper formulates perceptually inspired color enhancement through variational energy functionals, whose minimization exposes contrast enhancement and dispersion control. Tests support the predicted image effects, while parameter tuning and noise control remain open concerns.

9 Conclusions and future works

  • The authors translate human color vision phenomenology into assumptions on energy functionals and obtain color-correction algorithms by minimizing them.The minimization uses iterative gradient descent; the authors report stability and existence of minima.Source: §9 · pp. 24–25
  • The functionals’ minimization combines local nonlinear contrast enhancement with control of dispersion around middle gray and input values.The authors report that this framework makes the action of color correction on contrast and dispersion clearer.Source: §9 · p. 25
  • The functional class contains an energy leading to ACE and another corresponding to a kind of Retinex symmetrization.The paper identifies theoretical similarities and differences with both existing models.Source: §9 · p. 25
  • Tests on different images confirm theoretical predictions: the algorithms enhance details and remove color cast without introducing noise, though existing noise may be amplified.A pre- and post-processing procedure can be integrated to avoid noise amplification in very dark images.Source: §9 · p. 25
  • The algorithms use four explicit parameters, and image-dependent tuning that might improve enhancement remains beyond the paper’s scope and an open problem.A parameter set works satisfactorily across the test images.Source: §9 · p. 25
  • The authors identify improving noise control, including a functional treatment and study of noise versus local spatial frequency, as future work.Source: §9 · p. 25

Appendix

The appendix establishes existence and variation results for the proposed functionals, gives examples of absolute-value regularization, and analyzes conditions for stability and convergence of an iterative scheme.

9.1 The proof of Proposition 5.1

  • In the finite-dimensional image space, bounded minimizing sequences have convergent subsequences, and continuity of the energy terms yields a minimizer.Source: §9.1 · pp. 25–26

9.2 Examples of nice regularization of the absolute value

  • Examples of regularizing the absolute value have error bounds of order O(epsilon), including a bound uniform for z in [-1, 1].Source: §9.2 · p. 26

9.3 The proof of Proposition 5.2

  • For a symmetric pairwise functional with symmetric differentiable integrand, Lemma 9.1 expresses its first variation using derivatives with respect to the two arguments.The proof applies this result to regularized contrast functionals.Source: §9.3 · p. 27

9.4 Remarks on the stability of the algorithm

  • Under the stated condition on alpha, iterations stay in [rho, 1]; a further condition involving rho and m_epsilon gives convergence to a unique solution.The authors say that the sufficient convergence condition is impractical; with alpha = 255/253 and beta = 1, the algorithm exhibited stability and convergence in practice.Source: §9.4 · p. 29

Related papers

Loading 2511.23329v1…