A Computational Neuroscience Approach to Higher-Order Texture Perception




Seamons, John William George

Journal Title

Journal ISSN

Volume Title



Natural images contain large amounts of structural information characterised by higher-order spatial correlations. Neurons have limited capacities, so the visual system must filter out non-salient information, but retain that which is behaviourally relevant. Previous research has concentrated on two-point correlations; there has been less research into higher-order correlations, although the visual system is sensitive to them. Isotrigon textures can be used for this purpose. Their salient structure is exclusively due to fourth- and higher-order spatial correlations and they have the same structural features that create salience in natural images. In Chapter 2, we evaluated human texture discrimination using 10 novel isotrigon textures (VnL2) and 17 standard V3L2 isotrigon textures. Factor analysis revealed that as few as 3 mechanisms may govern the detection of fourth- and higher-order image structure. The Maddess group has previously published evidence that the number of independent mechanisms is less than 10 and perhaps as small as 3-4. The computation of higher-order correlations by the brain is neuro-physiologically plausible via nonlinear combinations of recursive and/or rectifying processes. In Chapter 3, we utilised the crowdsourcing platform “mTurk” to implement a large texture discrimination study. Under laboratory conditions, we showed that the testing modality was robust across a range of browsers, resolutions, contrasts and screen sizes. Texture discrimination data was gathered from 121 naïve subjects and compared to 2 independent laboratory data sets. Factor analysis indicated the presence of 3-4 factors, consistent with previous studies. Based on Pearson's correlation and coefficients of repeatability, mTurk is capable of producing data of comparable quality to laboratory studies. This is significant as mTurk has not previously been systematically evaluated for visual psychometric research. In Chapter 4, we employed a set of statistically controlled ternary textures. The textures were constrained (spatial correlations from 1st to 4th order) and their salience could be independently controlled by the addition of noise. To the ideal observer, all textures defined by a given amount of noise are equally detectable. However, humans are not ideal observers; their visual perceptual resources are restricted. Because of the number of textures available, we used mTurk to gather performance functions from 928 subjects for a subset of the texture space. Perceptual salience varied for each image statistic, with rank order: gamma > beta_hv > beta_diag > alpha > theta. This supports the order previously published for the related binary stochastic textures. The two least salient directions were consistently white:black and grey-bias (for gammas and betas), and black:grey and grey:white (for thetas and alphas). Such differences reflect the sensitivities and limitations of neural processing and are a manifestation of efficient coding. We hypothesised that the grey token conferred non-salience. Indeed, for gammas and betas, the grey-bias was consistently the second least salient. However, this did not hold for thetas or alphas. Counter-intuitively, the order of texture presentation did not significantly affect discrimination performance. An analysis of 31 repeat Workers found evidence of learning for beta textures, whereas performance for other textures was already maximal. This thesis concludes by considering future research.



texture, isotrigon, spatial correlation, crowdsourcing, computational neuroscience, perception, vision, visual, image statistics, computer vision, crowd sourcing




Thesis (PhD)

Book Title

Entity type

Access Statement

License Rights



Restricted until