Automatic Multilevel Feature Abstraction in Adaptable Machine Vision Systems

Vision is a complex task which can be accomplished with apparent ease by biological systems, but for which the design of artificial systems is difficult. Although machine vision systems can be successfully designed for a specific task, under certain conditions, they are likely to fail if circumstances change. This was the motivation for the research into ways in which systems can be self-designing and adaptable to new visual tasks. The research was conducted in three vital areas of concern for machine vision systems. The first area is finding a suitable architecture for forming an appropriate representation for the current task. The research investigated the application of Hypernetworks theory to building a multilevel, generally-applicable representation, through repeated application of a fundamental 'self-similarity' principle, that parts of objects assembled under a particular relation at one level, form whole objects at the next. Results show that this is potentially a powerful approach for autonomously generating an adaptable system-architecture suitable for multiple visual tasks. The second area is the autonomous extraction of suitable low-level features, which the research investigated through random generation of minimally-constrained pixel-configurations and algorithmic generation of homogeneous and heterogeneous polygons. The results suggest that, despite the simplicity of the features making them vulnerable to image transformations, these are promising approaches worth developing further. The third area is automatic feature selection. The research explored management of 'dimensionality' and of 'combinatorial explosion', as well as how to locate relevant features at multiple representation levels, in the context of 'emergence' of structure. Results indicate that this approach can find useful 'intermediate-level' constructs through analysis of the connectivity of the simplices representing objects at higher levels. The research concludes that the proposed novel approaches to tackling the above issues, in particular the application of hypernetworks to the formation of multilevel representations and the resulting emergence of higher-level structure, is fruitful