Textural Difference Enhancement based on Image Component Analysis
In this thesis, we propose a novel image enhancement method to magnify the textural differences in the images with respect to human visual characteristics. The method is intended to be a preprocessing step to improve the performance of the texture-based image segmentation algorithms. We propose to calculate the six Tamura's texture features (coarseness, contrast, directionality, line-likeness, regularity and roughness) in novel measurements. Each feature follows its original understanding of the certain texture characteristic, but is measured by some local low-level features, e.g., direction of the local edges, dynamic range of the local pixel intensities, kurtosis and skewness of the local image histogram. A discriminant texture feature selection method based on principal component analysis (PCA) is then proposed to find the most representative characteristics in describing textual differences in the image. We decompose the image into pairwise components representing the texture characteristics strongly and weakly, respectively. A set of wavelet-based soft thresholding methods are proposed as the dictionaries of morphological component analysis (MCA) to sparsely highlight the characteristics strongly and weakly from the image. The wavelet-based thresholding methods are proposed in pair, therefore each of the resulted pairwise components can exhibit one certain characteristic either strongly or weakly. We propose various wavelet-based manipulation methods to enhance the components separately. For each component representing a certain texture characteristic, a non-linear function is proposed to manipulate the wavelet coefficients of the component so that the component is enhanced with the corresponding characteristic accentuated independently while having little effect on other characteristics. Furthermore, the above three methods are combined into a uniform framework of image enhancement. Firstly, the texture characteristics differentiating different textures in the image are found. Secondly, the image is decomposed into components exhibiting these texture characteristics respectively. Thirdly, each component is manipulated to accentuate the corresponding texture characteristics exhibited there. After re-combining these manipulated components, the image is enhanced with the textural differences magnified with respect to the selected texture characteristics. The proposed textural differences enhancement method is used prior to both grayscale and colour image segmentation algorithms. The convincing results of improving the performance of different segmentation algorithms prove the potential of the proposed textural difference enhancement method.
Texture difference enhancement, Segmentation improvement, Human visual perception, Morphological component analysis, Wavelet-based dictionary
Doctor of Philosophy (Ph.D.)