Model Learning in Iconic Vision

Abstract

Generally, object recognition research falls into three main categories: (a) geometric, symbolic or structure based recognition, which is usually associated with CAD-based vision and 3-D object recognition; (b) property, vector or feature based recognition, involving techniques that vary from specific feature vectors, multiple filtering to global descriptors for shape, texture and colour; and (c) iconic or image based recognition, which either complies with the traditional sensor architecture of an uniform array of sampling units, or uses alternative representations. An example is the log-polar image, which is inspired by the human visual system and besides requiring less pixels, has some useful mathematical properties. The context of this thesis is a combination of the above categories in the sense that it investigates the area of iconic based recognition using image features and geometric relationships. It expands an existing vision system that operates by fixating at interesting regions in a scene, extracting a number of raw primal sketch features from a log-polar image and matching new regions to previously seen ones. Primal sketch features like edges, bars, blobs and ends are believed to take part of early visual processes in humans providing cues for an attention mechanism and more compact representations for the image data. In an earlier work, logic operators were defined to extract these features, but the results were not satisfactory. This thesis initially investigates the question of whether or not primal sketch features could be learned from log-polar images, and gives an affirmative answer. The feature extraction process was implemented using a neural network which learns examples of features in a window of receptive fields of the log-polar image. An architecture designed to encode the feature’s class, position, orientation and contrast has been proposed and tested. Success depended on the incorporation of a function that normalises the feature’s orientation and a PCA pre-processing module to produce better separation in the feature space. A strategy that combines synthetic and real features is used for the learning process. This thesis also provides an answer to the important, but so far not well explored, question of how to learn relationships from sets of iconic object models obtained from a set of images. An iconic model is defined as a set of regions, or object instances, that are similar to each other, organised into a geometric model specified by the relative scales, orientations, positions and similarity scores for each pair of image regions. Similarities are measured with a cross-correlation metric and relative scales and orientations are obtained from the best matched translational variants generated in the log-polar space. A solution to the structure learning problem is presented in terms of a graph based representation and algorithm. Vertices represent instances of an image neighbourhood found in the scenes. An edge in the graph represents a relationship between two neighbourhoods. Intra and inter model relationships are inferred by means of the cliques found in the graph, which leads to rigid geometric models inferred from the image evidence.