Abstract
The problem of scalable image recognition has long been a research issue in the area of computer vision. This thesis aims to address this by providing a holistic solution in which ensembles of very large number of classifiers for various image content are initially developed in order to cover as many perspectives of problem space as possible. The selection of these classifiers is then optimised simultaneously using evolutionary algorithms to ensure that the features and classifiers selected are optimal for the heterogeneous system components. To model image context, Hidden Markov Models are established through various evolutionary computation processes. Especially a hybrid evolutionary approach has been developed to find the most suitable contextual models which concurrently guide the searching for optimal classifiers. Finally information from optimised classifiers and context models are fused together to reason and determine the overall image content. This proposed architecture has been tested on Diabetic Retinopathy (DR) image datasets, which exhibit great variability and diversity. Based on the proposed solution, the system is able to recognise the key DR signs and ultimately, to separate normal and abnormal diabetic retinopathy images. Through evolutionary computation, the various components of the system outperformed classical approaches. This is demonstrated through the comparison between combined optimal classifiers and those obtained through traditional combination strategies such as average, sum and majority vote. Experiments also show that hybrid evolutionary approaches for optimising the classifier combination strategy and context models simultaneously perform best, if each component is treated individually. Using the integrated approach, the system developed is capable of separating normal and abnormal retina image with a Sensitivity of 95% and a Specificity of 92%. Among all images the system recognises as normal, 91% are true normal ones.