Optimizing image segmentation and classification methods in the presence of intensity heterogeneity and feature complexity

Abstract

Research in computer vision and image processing is expanding in various fields. This expansion creates the need to develop robust methods for acquiring qualitative and quantitative information from natural or real world images that can be used without or little customization in a diversity of applications. In natural scenes, intensity heterogeneity and feature complexity often occurs, thus fully automated computerized image processing and analysis methods are considered challenging processes. Image segmentation and classification are common processes for extracting valuable information from the analysed images. In image segmentation methods, the presence of intensity heterogeneity may lead to overlaps between the regions of interest and background objects or present artefacts. In image classification methods, intensity heterogeneity may lead to misperceptions of the various classes. The algorithms presented in this thesis, address image segmentation and classification in the presence of intensity heterogeneity and feature complexity, with applicability in mammographic and remote sensing image analysis. The presented algorithms extract quantitative and qualitative information from the analysed scenes acquired from the selected domains where the images often include intensity heterogeneity and feature complexity. The proposed segmentation algorithms are based on active contour and level set methodologies. Two different approaches are followed for designing optimal models for automatic segmentation. In the first approach, pre-processing methods are designed to optimize the overall performance of active contour segmentation models. Image transformations are produced and indices from these representations are utilized to form optimized active contour segmentation models. Intensity heterogeneity is eliminated or reduced by designing spectral and spatial filters. The sensitivity of filter responses to the parameter selection is addressed by applying an automated scheme for tuning and optimizing the filter settings. Furthermore, the method addresses the variation of colour intensity values that may be present in the analysed image and often leads to inaccurate results, by utilizing information from various colour systems. This information is analysed and used to selectively detect regions of interest that very often are not delineated by the traditional active contour models. Moreover, active contour models often tend to be sensitive on initialization. Thus, a clustering method is developed and incorporated in the segmentation algorithm to design optimal initial level set contours and drive the ii propagation process faster to achieve better segmentations. Additionally, statistical measures are used for designing optimum post-processing morphological filters to eliminate any misleading information that still exists in the final segmentation mask. In the second approach, the presence of intensity heterogeneity in complex scenes is addressed by designing an optimized localized active contour model. The proposed segmentation model is based on the decomposition of the image into intrinsic components and the use of a kernel convolution function for eliminating or reducing the intensity heterogeneity artefact. A new energy minimization active contour model for image segmentation where the presence of intensity heterogeneity is corrected during the propagation process of the active contours is suggested. In the classification domain, statistical distributions of various texture descriptors and their combination are investigated with support vector machines for the development of an objective image and/or regions of interest classification framework. Integrating various sets of features is fundamental for achieving better classification performance thus a model for the optimal integration of multiple feature sets for image classification in the presence of intensity heterogeneity is proposed. This thesis also suggests how the proposed methods may be combined in a single framework for the quantitative and qualitative image analysis in real world applications. The optimized active contour segmentation methods are used in mammographic images for automatic breast region delineation and abnormalities extraction. In satellite images, the proposed methods are applied for building boundaries and building shadows extraction. Subsequently, this information is utilized for estimating the height of the corresponding building structures. The classification framework is applied for breast density characterization and breast abnormalities detection. In addition, a no-reference image quality assessment model is developed based on the proposed classification model for extracting useful information about the quality of the images under evaluation. In all the proposed models, special attention is given on the automation of the implemented methods and their ability to deal with big data of various classes. Experimental results are shown at each developed stage of the thesis.