Biologically-inspired hierarchical architectures for object recognition

Abstract

The existing methods for machine vision translate the three-dimensional objects in the real world into two-dimensional images. These methods have achieved acceptable performances in recognising objects. However, the recognition performance drops dramatically when objects are transformed, for instance, the background, orientation, position in the image, and scale. The human’s visual cortex has evolved to form an efficient invariant representation of objects from within a scene. The superior performance of human can be explained by the feed-forward multi-layer hierarchical structure of human visual cortex, in addition to, the utilisation of different fields of vision depending on the recognition task. Therefore, the research community investigated building systems that mimic the hierarchical architecture of the human visual cortex as an ultimate objective. The aim of this thesis can be summarised as developing hierarchical models of the visual processing that tackle the remaining challenges of object recognition. To enhance the existing models of object recognition and to overcome the above-mentioned issues, three major contributions are made that can be summarised as the followings 1. building a hierarchical model within an abstract architecture that achieves good performances in challenging image object datasets; 2. investigating the contribution for each region of vision for object and scene images in order to increase the recognition performance and decrease the size of the processed data; 3. further enhance the performance of all existing models of object recognition by introducing hierarchical topologies that utilise the context in which the object is found to determine the identity of the object. Statement of