Image texture analysis using zero crossings information
G. Smith · The University of Queensland · 1998
This thesis describes a novel texture analysis algorithm. The algorithm is quantitatively compared with algorithms in the literature. The proposed algorithm is found to be more accurate than the existing algorithms. Further experiments indicate that the statistical peaking phenomenon, otherwise known as the curse of dimensionality, affects texture analysis experiments typical of those reported in the literature. Also, a taxonomy of stochastic models of texture is proposed. The history of definitions of texture in the literature is reviewed. The emerging consensus is that texture is best defined as a two-dimensional homogeneous random field. This thesis aims to describe a novel analysis algorithm which derives from this formal definition of texture, and whose features encode complete texture information. The novel texture algorithm uses a stochastic model of texture, and has its theoretical basis in the information present in zero-crossings of a signal. The completeness of the texture information encoded in the features can be analysed within the framework of zero-crossings theory. There is no widely accepted benchmark for texture segmentation algorithms. The novel algorithm is demonstrated on three collages of Brodatz images. Subjectively, the results are comparable to other segmentations in the literature. There is no widely accepted benchmark for texture classification algorithms. A subset of ten problems is selected from the MeasTex framework. The accuracy of the novel algorithm, and of several algorithms from the literature (Grey-Level Co-occurrence, Gaussian Markov Random Field, Gabor Energy and Ntuple algorithms), is measured on these problems. The novel algorithm compares favourably with the other algorithms on these problems, giving both the highest average accuracy and most consistent accuracy. It is demonstrated that the peaking phenomenon affects these experiments, which are typical of those reported in the literature. For all the algorithms investigated, the optimal parameter settings are influenced by the peaking phenomenon. With these parameter settings, none of the algorithms investigated encode complete texture information in their features.