Human colour vision is trichromatic: it is underpinned by the comparison of signals from three different classes of cone photoreceptor, those maximally sensitive to long (L), middle (M) and short (S) wavelengths. It can therefore be represented in a three-dimensional colour space in which the three axes represent the signals or combinations of signals from the three cone classes. Physiological data showing how these cone signals are combined and compared has motivated the creation of several important colour spaces, notably MacLeod-Boynton colour space, in which one axis of the space represents midget ganglion cells' comparison of L and M signals (Lee, Kremers & Yeh, 1998), while the other represents the small bistratified cells' comparison of S signals with summed L and M signals (Dacey & Lee, 1994). This space, along with other two- and three-dimensional colour spaces, have shaped the way researchers have imagined and investigated colour vision. However, while physiological measurements as far as the early cortex are consistent with linear combinations of cone inputs (Derrington, Krauskopf & Lennie, 1984; Solomon & Lennie, 2007; Sun, Smithson, Zaidi & Lee 2006), psychophysics has revealed a number of important effects that cannot be captured in any Euclidean colour space. Chapter 2 focuses on hue (the subjective quality of appearing, for example, blue, bluish-green or green) and saturation (the subjective counterpart of colorimetric purity, where a saturated colour can be desaturated to a pastel shade by adding white). The 'super-importance of hue' is a nonlinearity of colour vision in which a smaller physical stimulus change is required to detect a change in hue than to detect a change in saturation (Judd, 1968). Chapter 2 presents the first explicit test of Kuehni's 2003 claim that hue super-importance is the result of separate neural channels for hue and saturation, a claim which was ultimately not supported by the data. Further irregularities arise from individual differences in the cone mosaic that forms the gateway to the visual system. The cones of the three classes are not equal in number, and between individuals the ratio of L and M cones varies widely. Such differences are exacerbated locally when cones of the same class 'clump' together, which generally occurs to the degree that would be expected in a truly randomly distributed retina (Hofer, Carroll, Neitz, Neitz & Williams, 2005; Roorda, Metha, Lennie & 3 Williams, 2001). This should cause differences in spatiochromatic resolution within and between observers, but to date this has only been demonstrated in the fovea in a single observer with an extremely imbalanced L:M cone ratio (Miyahara, Pokorny, Smith, Baron & Baron, 1998) and in the parafovea (Danilova, Chan & Mollon, 2013). Chapter 3 reports the first evidence for an effect of L:M ratio on foveal spatiochromatic thresholds within and between observers who have more moderate L:M ratios. Until recently, information about the human cone mosaic could only come from histology, but the introduction of adaptive optics (AO) technology to retinal imaging has revolutionised our ability to investigate living, seeing human retinae at high resolution. This technology allows eye movements to be measured on an unprecedentedly fine scale referenced to the photoreceptor mosaic (Young, Hauperich, Morris, Saunter & Smithson, 2017), as described in Chapter 4. Chapter 5 describes how this capacity has been used to investigate a possible functional role for eye movements in chromatic vision. Due to the unequal numbers of cones of different classes and local clumping, fine chromatic detail that can be detected easily by one part of the retina may not be detected by another. Therefore, the hypothesis was tested that eye movements may help observers to overcome the irregularity of the retinal mosaic by translating the image across the retina and ensuring that the image is sampled by all three classes of cone. While visual inspection of certain trends was promising, no significant associations were found between particular amounts of eye movement and spatiochromatic performance, between eye movement and L:M ratio, or between eye movement and scene content. AO has more generally been used to extract important statistics about areas of retina, such as cone density, cone spacing and anisotropy in cone packing. However, the high volume of data generated requires automated methods of extraction, and such methods tend not to be readily generalisable between the images obtained from different labs' equipment. Chapter 6 describes the key prerequisites to improve the accuracy and automation of the Oxford Perception Lab's cone identification algorithm, comparing its outputs to carefully controlled manual cone counts. In summary, this thesis probes irregularities in the story of trichromatic colour vision: the possibility of separable neural channels that underlie specific nonlinearities of colour space, and the consequences and a possible mediator of the imbalanced numbers of L and M cones in the human retina. It additionally advances the development and assessment of processing methods for a technology, AO retinal imaging, that shows great promise for the study of colour vision in humans in vivo.