Image feature analysis using the Multiresolution Fourier Transform
The problem of identifying boundary contours or line structures is widely recognised as an important component in many applications of image analysis and computer vision. Typical solutions to the problem employ some form of edge detection followed by line following or, more commonly in recent years, Hough transforms. Because of the processing requirements of such methods and to try to improve the robustness of the algorithms, a number of authors have explored the use of multiresolution approaches to the problem. Non-parametric, iterative approaches such as relaxation labelling and "Snakes" have also been used. This thesis presents a boundary detection algorithm based on a multiresolution image representation, the Multiresolution Fourier Transform (MFT), which represents an image over a range of spatial/spatial-frequency resolutions. A quadtree based image model is described in which each leaf is a region which can be modelled using one of a set of feature classes. Consideration is given to using linear and circular arc features for this modelling, and frequency domain models are developed for them. A general model based decision process is presented and shown to be applicable to detecting local image features, selecting the most appropriate scale for modelling each region of the image and linking the local features into the region boundary structures of the image. The use of a consistent inference process for all of the subtasks used in the boundary detection represents a significant improvement over the adhoc assemblies of estimation and detection that have been common in previous work. Although the process is applied using a restricted set of local features, the framework presented allows for expansion of the number of boundary feature models and the possible inclusion of models of region properties. Results are presented demonstrating the effective application of these procedures to a number of synthetic and natural images.