Signal-linear representations of colour for computer vision
Type of content
Most cameras detect colour by using sensors that separate red, green and blue wavelengths of light which is similar to the human eye. As such most colour information available for computer vision is represented in this trichromatic colour model, Red Green Blue or RGB. However this colour model is inadequate for most applications as objects requiring analysis are subject to the reflective properties of light, causing RGB colour to change across object surfaces. Many colour models have been borrowed from other disciplines which transform the RGB colour space into dimensions which are decorrelated to the reflective properties of light.
Unfortunately signal noise is present in all acquired video, corrupting the image information. Fortunately most noise is statistically predictable, causing offsets from the true values following a Poisson distribution. When the standard deviation of a noise distribution is known, then noise can be stochastically predicted and accounted for.
However transformations inside cameras and transformations between colour models often deform the image information in ways that make the noise distributions non-uniform over the colour model. When computer vision applications need to account for non-uniform noise, wider tolerances are required overall. This results in a loss of useful information and a reduction in discriminative power.
This thesis has a focus on the linearity of signal noise distributions in colour representations which are decorrelated to the reflective properties of light. Existing colour models are described and each of their components examined with their strengths and weaknesses discussed.
The results show that the proposed Signal Linear RGB (SLRGB) colour model achieves a transformation of the RGB colour space with uniform noise distributions along all axes under changes to camera properties. This colour space maintains a signal noise with a standard deviation of one unit across the space under changes of the camera parameters: white balance, exposure and gain. Experiments demonstrated that this proposed SLRGB model consistently provided improvements to linearity over RGB when used as a basis for other colour models.
The proposed Minimum Weighted Colour Comparison (MWCC) method allows reflectively decorrelated colour models to make colour comparisons which counter the deforming effects of their coordinate systems. This was shown to provide substantial improvements to linearity tests in every case, making many colour models have a comparative noise linearity to undeformed colour models.
The proposed Planar Hue Luminance Saturation (PHLS) and Spherical Hue Luminance Saturation (SHLS) colour models are decorrelated to reflective properties of light and allow for signal linear colour comparisons. When used for pixel classification of coloured objects the PHLS and SHLS colour models used only 0.26% and 0.25% of the colour volume to classify all of the objects, with the next best using 0.88% without MWCC and 0.45% with.
The proposed Gamut Limit Invariant (GLI) colour model extends the decorrelation of reflective properties of light further by correcting for colours which are too bright and are clipped by the limits of the RGB space. When clipping occurs the properties become no longer decorrelated and shift. GLI models these changes to estimate the original values for clipped colours. The results show that this method improves decorrelation when performing pixel classification of coloured objects with varying proportions of clipped colours.
Overall, the results show that the proposed framework of colour models and methods are a significant improvement over all prior colour models in enabling the most accurate information possible for processing colour images.