Novel textural features and techniques for image segmentation and classification

Texture analysis is an important tool in pattern recognition and image processing. Humans can recognize and classify different textural regions quite accurately, but they cannot analyze, categorize or process a large amount of data fast. The objective of this dissertation is the development of feature extraction, clustering and classification schemes in order to simulate the human ability to recognize and categorize textures.; Chapter 3 deals with the development of features, which are independent to image intensity and contrast variations. A multiple fractal dimension-based feature extraction scheme is proposed. The fractal-based set is compared to the traditional energy/Gabor filter bank decomposition set. Results show that the fractal set identifies different uniform textural regions without getting affected by local intensity and contrast variations. An iterative K-means clustering technique has been used for segmentation that has been proven effective in determining the texture boundaries compared to standard K-means. Also, an index for the determining the number of different textural regions is examined.; Fractal dimension is constructed from all available scales. In chapter 4, we develop a wavelet-based approach for computation of fractal-like, single scale features for a multi-resolution texture description. Feature comparison in terms of percentage of correct classification for a large database illustrated the improved performance of the proposed set.; An important human characteristic is the ability to recognize objects irrespective of the observation rotation angle and at the same time retain the texture directionality component. In chapter 4, wavelet-based roughness features are introduced. Segmentation and classification results show that these features satisfy the aforementioned requirements.; In chapter 5, a neural network classifier is employed to classify textures affected by noise. The algorithm is a modification of the Fuzzy ARTMAP neural network that exhibits superior classification performance than the standard Fuzzy ARTMAP.; In chapter 6, textural features in combination to intensity measures are used for the removal of unwanted echoes from weather radars. Comparisons with an existing manual algorithm verify that one measure (reflectivity) is sufficient for a good quality control algorithm. Other algorithms require knowledge of meteorological parameters such as pressure and humidity or other measures (radial velocity, spectrum width).