SPATIAL FILTERING IN MULTI-RESOLUTION TEXTURE IMAGE ANALYSIS (DIGITAL IMAGE PROCESSING).

摘要

Available from UMI in association with The British Library. Requires signed TDF.; In digital image processing, texture has always been considered a vital element for both modelling and analysis of images. One method of analysing texture is to examine it at a number of different resolutions. The thesis focuses on issues arising from this multi-resolution aspect of texture.; The lack of a standard database of digitised pictures for texture analysis has meant that individual investigators have to generate their own, resulting in lack of comparability between different versions of the same texture. Since multi-resolution texture analysis is dependant on the distribution of grey level statistics of the digitised images, factors which affect the digitisation procedure are briefly discussed.; Most of the texture analysis algorithms produce a number of images at different resolutions from the original image. An interpretation of this approach as a 'measurement space' is developed in the thesis. The measurement space contains a number of measurement planes each one corresponding to a specific resolution (or scale). As an example, the thesis defines the measurement space and its corresponding measurement planes for a range of known multi-resolution approaches.; Since multi-resolution features are defined and extracted on measurement planes, it is possible that interaction between adjacent planes in the same measurement space might occur. This interaction, which appears not to have received the proper attention in the literature, is the key issue of the research described in the thesis. It can introduce an effect analogous to aliasing which may seriously impair the performance of existing multi-resolution texture analysis and classification algorithms. It is shown experimentally that this interaction is not important in the case of textures following a presumed 'fractal' model. It is also shown that images of natural textures do not follow generally the fractal model, in contrast to previously reported results. A brief presentation of the theory of fractal geometry precedes the presentation of simulation experiments. The thesis also shows that the performance of two existing multi-resolution algorithms can be improved if interaction between scales is eliminated after appropriate filtering.