Morphological granulometry for texture analysis

This thesis concerns the analysis of digital texture images, using techniques from mathematical morphology and regression modelling for the classification of texture images. It investigates the use of granulometric moments, arising from the morphological pattern spectrum, as texture descriptors to predict evolution time or class label of texture images which evolve over time or follow an intrinsic ordering of textures. A cubic polynomial regression is used to model each of several granulometric moments as a function of time or class. These models are combined in a novel way and used to predict time or class. The methodology was developed on synthetic images of evolving textures generated for the purpose, and then applied to classify a sequence of images of corroding metal to a point on an evolution time scale. Performance of the new regression approach is compared to that of several well established classifiers, namely linear discriminant analysis, neural networks and support vector machines (SVMs). The method was also applied to images of Indian black tea granules, which are ordered according to granule size. Better classification was achieved for both sets of images compared to previously published results for these images. The performance of grey level co-occurrence matrix (GLCM) features from the synthetic images and both sets of real images was compared to that of granulometric moments, and it was found that granulometric moments provide much improved classification compared to GLCM features for such shape-based texture images. The performance of wavelet-based features from the Indian black tea images was also evaluated and was poorer than expected. SVMs were generally found to be superior to the other classifiers. The later part of this thesis concerns classifying hyperspectral images of Chinese teas. Several methods were compared for selection of appropriate spectral bands from these images. Principal component analysis and entropy proved to be the best band selection criteria in this application. GLCM features, wavelet-based features and wavelet-based GLCM features outperformed granulometric moments computed from the same set of bands. Calculating texture features from an optimum set of spectral bands gave better classification performance compared to the use of RGB (red, green and blue) or HSV (hue, saturation and value) colour representations or grey scale versions of the images.

Name dc:type.qualificationname

phd

Level dc:type.qualificationlevel

doctoral-pg

Grantor dc:publisher.institution

University of Strathclyde

Year dc:date.issued

2012

Author dc:creator

• Khatun, Mahmuda

DOI

10.48730/nh81-h884

Identifier

T13327

OAI identifier oai:identifier

oai:strathclyde:xg94hp534