Feature-Based Segmentation Of Textured Images

The segmentation of textured images aims to partition an image into severaldisjointed regions that are homogeneous with regards to some texture measures, sothat subsequent higher level computer vision processing can be performed. It has longbeen one of the most important branches of digital image processing and has drawnconsiderable attention of researchers from around the world. During the past threedecades, hundreds of segmentation algorithms have been proposed in the literature.However, due to the diversity of images, the complexity of natural textures and thelack of understanding of the human vision system (HVS), those algorithms usuallysuffer from less accuracy and narrow image specific orientation. Therefore, texturesegmentation is, up to now, still an open topic with great challenge in imageprocessing field.This dissertation is devoted to the semi-supervised segmentation of textured graylevel images, where the number of texture patterns is known but the information abouttheir properties is not. After comprehensively reviewing the basic principles andexisted methods, the author chooses the feature-based approaches to solve thisproblem. Generally, feature-based texture segmentation algorithms can be viewed asconsisting of two successive processes: feature extraction and feature partition.Feature extraction tends to find an appropriate descriptor to characterize thehomogeneity of each texture in an image so that all pixels from the same texture canbe represented by vectors of similar value. Feature partition is a process of assigningeach feature a label to designate the region or class to which it belongs, and thussegmentation result can be obtained through the relationship between features andpixels. In this dissertation, the author investigates those two processes, respectively,and achieves highly effective, increasingly innovative and cutting-edge approaches oftexture segmentation, which can be summarized as follows.1. An overview of segmentation of textured images, including the fundamentaldefinitions, the research background and the significance of this topic is presentedand the mainstream approaches and the state of art of algorithms in this field arereviewed in this dissertation.2. Texture feature extraction is investigated by using the fractal model in this dissertation. Various fractal dimensions have been widely used as texturedescriptors. However, the popular box-counting based fractal dimension iscommonly criticized for its less accuracy,, which is mainly caused by the regularpartition and counting scheme. Through analyzing the disadvantages of thetraditional morphological method, the author proposes a modified morphologicalfractal estimation approach, which uses a series of structure elements withdifferent scales to take the place of the unit structure elements used by traditionalmethod so that the estimation accuracy has been further improved. Throughdelicately selecting the shape of structure elements and constructing an iterativedilation scheme, the proposed approach substantially reduces the computationaltime-cost. When applied to texture segmentation, the novel morphological fractaldimension demonstrates an improved ability to differentiate various textures.3. Texture features based on the multifractal model is studied in this dissertation.Due to the limited bit depth and spatial resolution, most digital images are merelysemi-fractals and have anisotropic and inhomogeneous scaling properties.Therefore, fractal dimension alone is intrinsically not sufficient to representtexture patterns. To characterize the fractal reality of textured images, the authorgeneralizes the morphological fractal estimation algorithm to multifractalestimation, and thus proposes a novel texture descriptor called the localmorphological multifractal exponents (LMME). Furthermore, motivated by theidea of fractal signature, the author extends the LMME feature to themorphological multifractal signatures (MMFS). Those two multifractal texturefeatures has been compared with other commonly uses features in segmentation oftexture mosaics. The experimental results demonstrate that the novel features candifferentiate textured images more effectively and provide more robustsegmentations.4. Feature partition based on fuzzy clustering is explored in this dissertation. Featurepartition in feature-based texture segmentation is different from traditional patternclassification problems in that texture features imply not only the position infeature space but the position on image surface. Therefore, a texture feature isindeed a spatial pattern so that a textured image can be modeled as a set of spatialpatterns. The author proposes an approach to perceptual segmentation of imagesthrough the means of fuzzy clustering of spatial patterns, where the distancebetween a spatial pattern and each cluster is defined as a combination of the Euclidean distance in the feature space and the spatial dissimilarity, which reflectshow much of the pattern's neighbourhood is occupied by other clusters. Theresults of comparative experiments demonstrate that the proposed approach cansubstantially improve the segmentation accuracy. Moreover, the author alsogeneralizes this approach to a multi-level feature partition algorithm, whichsignificantly decreases the computational complexity of texture segmentation.5. In feature-based texture segmentation, feature estimation and feature partition arenot two independent processes. Regardless of this fact, traditional methods usuallysuffer from the less accuracy, which is intrinsically caused by the oversimplifiedassumption that each textured sub-image used to estimate a feature ishomogeneous. To solve this problem, the author proposes a coupled Markovrandom field (CMRF) model, which has two coupled components: one models theobserved image to estimate features, and the other models the labeling to achievefeature partition. When calculating the feature of each pixel, the homogeneity ofthe sub-image is ensured by using only the pixels currently labeled as the samepattern. With the acquired features, and the labeling is obtained through solving aMAP (maximum a posteriori) problem. In this adaptive segmentation approach,the features and the labeling are mutually dependent on each other, and thereforeare alternately optimized by a simulated annealing scheme. With the gradualimprovement of features' accuracy, the labeling is able to locate the exactboundary of each texture pattern. The proposed algorithm is compared with asimple MRF model based method in segmentation of both Brodatz texturemosaics and real scene images. The satisfying experimental results demonstratethat the proposed approach can differentiate textured images more accurately.