Evolving color constancy

https://doi.org/10.1016/j.patrec.2005.07.020Get rights and content

Abstract

The ability to compute color constant descriptors of objects in view irrespective of the light illuminating the scene is called color constancy. We have used genetic programming to evolve an algorithm for color constancy. The algorithm runs on a grid of processing elements. Each processing element is connected to neighboring processing elements. Information exchange can therefore only occur locally. Randomly generated color Mondrians were used as test cases. The evolved individual was tested on synthetic as well as real input images. Encouraged by these results we developed a parallel algorithm for color constancy. This algorithm is based on the computation of local space average color. Local space average color is used to estimate the illuminant locally for each image pixel. Given an estimate of the illuminant, we can compute the reflectances of the corresponding object points. The algorithm can be easily mapped to a neural architecture and could be implemented directly in CCD or CMOS chips used in todays cameras.

Section snippets

Motivation

The human visual system is able to determine the color of objects irrespective of the color of the illuminant (Land, 1964, Land, 1974, Zeki, 1993). If a room with a white wall is illuminated with a yellow light, i.e. caused by a yellow lamp shade, the wall nevertheless appears white to a human observer. However, if the observer takes a photograph of the room then the wall will appear yellow in the photograph. The human visual system is somehow able to compute color constant descriptors which do

Theory of color image formation

We now briefly review some background material on color image formation. For this, we consider a planar patch located at some distance in front of a camera or measuring device. We assume that the measuring device has a number of different sensors which respond to light in different parts of the spectrum. The human eye contains three types of cones which respond mainly to the light in the red, green and blue part of the spectrum. Similarly, a digital camera contains light sensors with red,

Evolving an algorithm for color constancy

Our goal was to find a biologically plausible algorithm, i.e. an algorithm which could be mapped to a neural architecture, for color constancy. Note that most existing color constancy algorithms, apart from the approaches using neural nets (Cardei and Funt, 1999, Courtney et al., 1995, Dufort and Lumsden, 1991, Funt et al., 1996, Pomierski and Groß, 1995, Usui and Nakauchi, 1997), are quite complicated. Since we are turning to artificial evolution for an answer to the problem color constancy,

A parallel algorithm for color constancy

Encouraged by the above results, we developed a parallel algorithm for color constancy (Ebner, 2002, Ebner, 2004). As above, we assume that we have a grid of processing elements, one processing element per pixel. The algorithm is based on the computation of local space average color. Let us assume that the reflectances of the objects are evenly distributed in the range [0, 1] and that a single uniform illuminant is used. Let us also assume that the viewed scene is sufficiently diverse, i.e.

Comparison with other algorithms for color constancy

Numerous algorithms for color constancy have been developed. However, most algorithms would be hard to implement using a neural architecture. They are not biologically plausible. A notable exception is the retinex algorithm developed by Land and McCann (1974) which was later extended to two dimensions by Horn (1974) and refined by Blake (1985). Horn suggests to first take the logarithm of the input intensity. This separates the product of reflectance and illumination into a sum.logci=logRi+logLi

Evaluation of algorithms on an object recognition task

We have used color based object recognition to evaluate the performance of the algorithms on the datasets of Barnard et al. (2002). Five different sets were used from this database. The images were down-sampled to 50% of the original size in order to speed up the evaluation. Image set 1 contains mainly Lambertian reflectors. Image set 2 contains objects with metallic specularities. Image set 3 contains objects with non-negligible dielectric specularities. Image set 4 contains objects with at

Conclusion

Color constancy is an important problem in many areas of computer vision research. In particular, color constancy is required to develop robust algorithms for service robots which have to work under changing lighting conditions. It is equally important for consumer photography. In order to develop algorithms which mimic the algorithms used by the human visual system we also have to look at the neural architecture of the brain. The whole visual system is a product of natural evolution and we

Acknowledgement

We have used the lilgp Programming System Vers. 1.1 (Zongker and Punch, 1996), for our experiments.

References (58)

  • Cardei, V.C., Funt, B., 1999. Committee-based color constancy. In: Proc. of the IS&T/SID 7th Color Imaging Conference:...
  • S.M. Courtney et al.

    A multistage neural network for color constancy and color induction

    IEEE Trans. Neural Networks

    (1995)
  • P.A. Dufort et al.

    Color categorization and color constancy in a neural network model of V4

    Biol. Cybernet.

    (1991)
  • M. D’Zmura et al.

    Mechanisms of color constancy

  • M. Ebner

    Evolving color constancy for an artificial retina

  • Ebner, M., 2002. A parallel algorithm for color constancy. Tech. Report 296, Universität Würzburg, Lehrstuhl für...
  • M. Ebner

    Color constancy using local color shifts

  • G.D. Finlayson

    Color in perspective

    IEEE Trans. Pattern Anal. Machine Intell.

    (1996)
  • G.D. Finlayson et al.

    Color constancy: generalized diagonal transforms suffice

    J. Opt. Soc. Amer. A

    (1994)
  • G.D. Finlayson et al.

    Spectral sharpening: sensor transformations for improved color constancy

    J. Opt. Soc. Amer. A

    (1994)
  • G.D. Finlayson et al.

    Intrinsic images by entropy minimization

  • Finlayson, G.D., Hubel, P.M., Hordley, S., 1997. Color by correlation. In: Proc. of IS&T/SID. The 5th Color Imaging...
  • G.D. Finlayson et al.

    Solving for colour constancy using a constrained dichromatic reflection model

    Interat. J. Comput. Vision

    (2001)
  • G.D. Finlayson et al.

    Comprehensive colour image normalization

  • D.A. Forsyth

    A novel approach to colour constancy

  • D.A. Forsyth

    A novel algorithm for color constancy

  • B. Funt et al.

    Is machine colour constancy good enough?

  • Funt, B., Cardei, V., Barnard, K., 1996. Learning color constancy. In: Proc. of the IS&T/SID 4th Color Imaging...
  • B.V. Funt et al.

    Color constancy computation in near-Mondrian scenes using a finite dimensional linear model

  • Cited by (33)

    • Multi-objective optimization based color constancy

      2014, Applied Soft Computing Journal
      Citation Excerpt :

      We have achieved the goal of the first phase in the proposed approach (which is high performance in classification and determination of the best group of algorithms for each image); but certainly, there may be other combinations that result in the same (or slightly better) classification performance. In the case of quantization step in edge direction histogram, the initial value has been chosen as 10 degree which has been suggested in [18] and after mentioned process in previous paragraph, we have found that each value in range [5,10] results in the same performance (and better than other examined values) while slightly changing the optimum for dimensionality reduction parameter. The values for σ in computation of Weibull parameters have also been chosen based on NIS method [17] which its source code is available for download in [27].

    • COLOR CONSTANCY BEYOND STANDARD ILLUMINANTS

      2022, Proceedings - International Conference on Image Processing, ICIP
    • The customized PCB colorchecker: When color restoration meets machine learning

      2019, ACM International Conference Proceeding Series
    • Illuminant chromaticity estimation via optimization of RGB channel standard deviation

      2016, Final Program and Proceedings - IS and T/SID Color Imaging Conference
    • Vision and evolution: State of the art

      2016, Natural Computing Series
    View all citing articles on Scopus
    View full text