Title of Thesis
Signal Processing Methods for Texture
Classification: Analysis and Improvements
Department of Computer and Information Sciences
/ Pakistan Institute of Engineering & Applied Sciences, Islamabad
|Number of Pages|
|Keywords (Extracted from title, table of contents and
abstract of thesis)|
Signal, Processing, Methods,
Texture, Analysis, Classification, Analysis, Improvements, Gabor,
This thesis investigates the signal processing methods for texture classification. Most of these methods involve comparatively uncomplicated and hardware viable image analysis techniques. Literature survey shows that for last several decades, these methods have been among the most popular tools for texture recognition tasks.
A comparative study of these methods from the dissimilarity analysis point of view has been performed. Improvement in discrimination capability of texture feature extraction techniques using edge operators and laws filtered images, through contrast stretching, have been proposed as well. Subsequently, fusion of the proposed texture features is studied and the fused set has been investigated from feature selection point of view.
The performance of texture features, extracted for segmentation using statistical methods, depends on an optimum region size called window. In most of the reported work, window-size selection is generally done through visual inspection or trial and error based techniques. The issue of optimum window for moment based texture feature extraction has been investigated and a framework based on Fourier analysis is formulated in order to automate the optimum window size computation and feature weight selection.
Gabor feature extraction is an established technique in describing the texture having features in the range of low frequencies. However, in the presence of periodic variance or impulsive noise, Gabor filters generate variable features at high frequencies. The fusion of optimized moment and Gabor energy texture features has been proposed as well.
In machine vision, rotation invariant feature extraction is one of the most challenging texture analysis tasks. This issue is addressed by proposing a novel moment invariant based feature set for efficient texture segmentation. In deriving proposed feature set, a moment mask based technique has been employed innovatively and in the process only seven moment images and computed.