Colour constancy algorithms differ in their derivation, implementation, performance and assumptions. The focus of the research presented in this thesis is to discover colour constancy solutions to recover surface colours, or equivalently, to estimate the illumination, of single light source in a given scene. Several colour constancy models will be proposed. These methods have different methodologies and constraints. For example, a method can be constrained on a particular model surface material, on blackbody radiation light source, on dichromatic model, and on spatial variation of the illumination and the reflectance. The methods to be discussed include, for instance, a method of identifying achromatic surfaces, which can then be used as known references for estimating the scene illumination. A second method examines the colour of human skin and its dependence on its hemoglobin content, melanin content, and the illuminating light. The corresponding basis of these three factors can be represented linearly in logarithm space, where the colour of the light can then be estimated. A third method, uses the fact that the colours reflected by an inhomogeneous dielectric material lie on a plane spanned by the colour of the specular component reflected from the air-surface interface and the colour reflected from the body of the material. Once these planes are detected by a Hough transform, their intersection line represents the scene illumination. A fourth method is based on the independence and difference in the rate of spatial variation of the luminance and the surface reflectance in a given scene, from which image features can be separated via non-negative matrix factorization to reveal the true surface reflectance. A fifth method is based on learning the correspondence between an image’s colour content and its illumination via thin-plate-spline interpolation so that the chromaticity of the light can be calculated. Finally, a quaternion-based curvature measure approach is developed that can be used as a complement to colour constancy methods that use information from spatial edges. In this thesis, these various methods are proposed to overcome drawbacks in existing approaches for better performance and improved robustness and efficiency.
Copyright is held by the author.
Member of collection