Texture direction analysis and segmentation using transition edges
Description
Texture can be defined as a structure composed of a large number of more or less ordered similar elements or patterns. Texture can be modelled as an arrangement of visible 'edges' which in turn amalgamate into patterns. This dissertation is concerned with the analysis of the angular orientation of the textural surfaces. Using the extracted texture features, segmentation of a multi-textural scene in a single, static image without any knowledge of the number of different texture regions or the types of texture is also developed. The images must be well defined and sharp in contrast for a good performance of the models. The images must be thresholded properly to bring out the predominant features of the texture. The directional analysis is achieved by counting the total number of edge transitions in four different directions viz., horizontal, vertical, right diagonal and left diagonal directions. A model relating the four directional edge transition counts is developed and used to determine the predominant orientation of the pattern. This technique provides a quick and efficient way to segment complicated images like multiple wood grains with different surface orientation. The directionality feature or the total of the horizontal and vertical edges or a combination of both can then be used as texture parameters to classify the texture scene. The proposed segmentation procedure uses a region-based technique. An optimal window size chosen from a set of square windows whose linear dimensions are powers of two is defined by minimizing an error function which includes the fraction of the image area misclassified by the texture segmenting procedure. A probability-based model has been developed to estimate the optimal window size. Another mathematical model to select the optimum window size is also developed and presented The results of the mathematical model are compared to that of the probability-based model. The optimum window selection will essentially provide an efficient window size to successfully segment the whole image into homogeneous regions. A complementary fine tuning procedure to increase the accuracy is proposed. This fine tuning procedure uses an edge-based approach by taking the gray level statistics into account. These techniques are shown to be effective with a wide variety of artificial and natural textures, both regular and random