A natural vision inspired approach to line and curve pattern analysis

Abstract

It is demonstrated that the Local Energy at each image location is rich in information where a variety of new and useful pattern analysis algorithms may be applied. In particular, several different types of line features may be detected and extracted by analysing the energy information in various orientation and frequency channels. The different types of line features are building blocks to form more complex patterns giving rise to a particular texture, such as text. Local energy was first proposed (Morrone and Burr 1988) as a generalised model for feature detection in human vision and has been shown to work well in texture analysis. This thesis focuses on developing biologically plausible algorithms of various pre-attentive feature detection tasks such as straight-line and curved-line feature discrimination and curvature discrimination. In biological vision, the term pre-attentive is used to describe visual tasks that may be performed without the need to engage the attention mechanism, or high level processing. The straight-line and curved-line feature discrimination algorithm involves filtering the input image with a set of directional filters covering the entire image space and computing the filter response energy. Features are extracted by analysing the energy information in various orientation channels where a straight-line will generate energy in a single orientation channel. In contrast, the energy generated by a curved-line feature is spread over multiple adjacent neighbouring channels. In addition to the discrimination of straightlines and curved-lines, the curvature of a curved-line feature can be obtained by analysing the energy distribution along the orientation dimension. Furthermore, the detection algorithm can segregate open and closed curved-line (e.g. circles) features. This thesis also presents a biologically inspired texture-based algorithm using the energy approach for (1) segmenting text embedded in clutter and (2) classifying text scripts without any explicit knowledge of the type of text present. The algorithm for text segmentation utilises the energy information in both the orientation and frequency dimensions. Similarly, energy analysis is applied to the task of text script classification using a set of descriptors derived from the energy information. The resulting algorithms are invariant to scale, rotation and position changes of text and are robust under noisy conditions. The algorithms detailed in this thesis offer a suitable biologically plausible solution in computer vision for various pattern analysis tasks such as line feature detection and classification. Finally it is demonstrated that a massively parallel implementation of the algorithms using Artificial Neural Networks is possible.