Data representation and models beyond human imitation

Image classification is a rapidly progressing field of computer vision, dominated by deep convolutional neural networks(DCNNs) that focuses on predicting labels based on input images. Current preprocessing techniques do not exploit the potential for data transformations to improve class separation despite the potential for these data transformations to make problems easier. Class separation is hugely relevant to classification difficulty, including being a major factor in the impact of dataset imbalance. Past projects and other research show that colour spaces transformations, a data transformation technique manipulating the colour spectra representation, can have value on classification difficulty including in imbalanced learning situations. This work however, both on imbalance problems and beyond, is vastly underdeveloped as it works within an existing canon of colour spaces, only a small fraction of which have been developed for computer vision tasks, with most being developed before computer vision was even conceived of. In this research the value of new spectral representations beyond imitations of human physiology or practically focused tools will motivate further examination of the potential for data transformation and aim to develop techniques to exploit their ability to improve class separation.