A connectionist approach to visual recognition

Strictly sequential approaches to computer vision are at best slow and cumbersome, at worst impossible. In this thesis, a massively parallel, connectionist approach is brought to bear on the problem of visual recognition. Computing with connections is a synthesis of results from Neuroscience, Computer Science, and Psychology. The fundamental assumption of connectionism is that individual computing units do not transmit large amounts of symbolic information. Instead, these units compute by being appropriately connected in a network of similar units. Using the communication pathways (connections) defined by the arcs of the network, the units cooperate and compete towards a globally consistent interpretation of the input scene. The problem, visual recognition, is defined as matching instances of predefined objects in the input with a fixed set of internal models. Predefined objects come from Kanade's Origami World{Kanade78}. The program represents and recognizes such pre-defined objects from line drawing input. To organize these networks, conceptual hierarchies are defined. A conceptual hierarchy is a semantic network hierarchically arranged according to abstraction levels. Levels represent the extraction of progressively more complex features. A node on a level, a computing unit, represents an instantiation of a feature defined on that level. Connections represent the composition and competition relations between feature units. Iterative relaxation is the form of control in the network. Each unit iteratively computes activation levels, a reflection of current confidence in the associated feature. Numerous testcases illustrate network behavior in presence of perfect, noised, incomplete and occluded input. The greatest benefits of this approach are seen in the systems ability to cope with incomplete and occluded input. Another advantage is an inherently parallel approach that eliminates the need for establishing a search order on exploring interpretations. The system exhibits many similarities with the structure and behavior of animal vision systems.