Exploring the visual pathway and its applications to image reconstruction, contrast enhancement and object recognition

The natural world is filled with perfectly working, functional systems that are robust, accurate, and adaptable; this work takes favour with the aforesaid and presents a biologically inspired approach to computer vision; in particular on the subjects of image reconstruction, contrast enhancement and object recognition. The first half of this thesis takes an exploratory approach, on the example of image reconstruction, towards the understanding of the visual pathway from retina to the primary visual cortex (V1), investigating redundancy reduction, information preservation and contrast enhancement. The retina having approximately 130 million cells, is forced to discriminate with the incoming information. Programmed for concision, the primate eye encodes information with sparsity, yet remains information preserving by encoding only contrast. By reconstructing an image from its contrast map pairs using gradient descent least squares error minimization, this work has shown that information is preserved across the optic nerve channel despite sparsification of the input image presented on the photoreceptors. By mimicking the irregularities of the eye's receptive fields, it has been shown that the neural architecture along the visual pathway is robust and fault tolerant against irregularities - a general characteristic of the entire nervous system. Using non-linear and asymmetric gain control with the on-and off-centre contrast map pairs, it has been shown that the mean luminance of an image can be controlled and the aforesaid reconstruction can be used for straightforward enhancement; thus reducing contrast enhancement to a scaling operation over the contrast domain. This has further been successfully applied to colour image contrast enhancement using a number of different models, including the neuro-physiologically proven representation of colour opponency, in the form of colour opponent contrast maps. With the above work serving as a pre-processing stage, the second half of the thesis approaches the subject of object recognition; improving upon prior work in Sparse Representation Classification (SRC). Sparseness is a key feature of the brain's internal representation whereby it achieves its robustness and adaptability. This work replaces the mean square error measure for similarity comparison of images with a perceptually compatible structural error measure, as well as the conventional sparsifiers with a genetic algorithm of the original SRC algorithm. This has resulted in an improved recognition rate - owed in large part to a more effective similarity comparison and improved sparseness of the solution. The approaches to the troika of reconstruction, contrast enhancement and object recognition strengthen both premise and belief that biologically inspired vision is dually meritorious and warrants greater appreciation and study by the Computer Vision community at large; not to be discounted as is often done. The hope is this work proves a seed for future endeavours