Research On Application Of Texture Feature To Image Segmentation

The segmentation of textured images has been one of the focuses and difficulties in the computer vision field all the time. Due to the complexity and diversity of natural textures, and the lack of understanding of human vision system, there is still a huge gap between the segmentation effect by machine and that by human. However, texture segmentation is the foundation of some high-level visual processing such as object recognition and image understanding, so it is always an important research topic.Image segmentation aims to separate the regions with different statistical characteristics in image. Countless segmentation algorithms have developed in the past decades, but each algorithm has its limitations and deficiencies. Texture is an important cue for segmentation. Generally, the texture feature based segmentation algorithm consists of two successive steps:feature extraction and image segmentation. Feature extraction is vital, from which the subsequent segmentation is strongly dependent. Feature extraction is a process describing the essential attribute of textured images and mapping all pixels of the same texture into similar vectors.Inspired by the processing of biological vision, this thesis uses a "filter-rectify-filter" model to capture the second-order features of texture since texture belongs to the second-order stimuli. Where, the parameters of filters are determined by the experimental results of visual psychology. The advantage is that the model conforms to human visual perception mechanism and can represent the essential attribute of texture. This method is a new way for solving the problems on textured image segmentation. In addition, this thesis dinamically introduces space information during clustering in feature space so that some small misclassified regions can be corrected. Experiments on both synthetic and real images showed that the segmentation based on the combination of second-order features and spatial information can obtain better results.From the perspective of mathematics modeling, the author has investigated texture segmentation algorithms based on amplitude-modulation frequency-modulation models. In order to eliminate the bad effect on subsequent segmentation which is due to the high dimensionality of the feature vector, traditional methods select parameters from an optimal channel to extract features by dominant component analysis. However, this approach inevitably brings the problem of information loss. This thesis elaborately selects and combines the features according to actual physical meaning, the obtained feature vector has lower dimensionality and remains enough useful information. Since the feature vector reflects the essential attribute of texture and presents a ball-like shape in feature space, a simple K-means clustering algorithm is utilized to obtain satisfactory segmentation results in next step. To verify the validity of the proposed method, experiments are performed on several open segmentation databases, and the quantitative comparison with existed algorithms based on modulation models is reported. Finally, the method is applied to remote sensing image segmentation.Texture segmentation is a very complex task. To describe textured images in detail, hierarchical structure is considered to express the image from coarseness to fine in different resolution. A new hierarchical segmentation mechanism is proposed in this thesis, the number of levels and the number of segments at each level are implicitly determined by user interaction. Therefore, the segmentation result at each level has its clear semantics.User could group these results according to resolution request of the high-level visual task. At each level of the tree map, a maximal similarity based region merging algorithm is used to implement segmentation task, where, the similarity is measured by the L*a*b* color feature and Gabor energy feature of two adjacent regions. Experiments demonstrate that, the proposed method is superior to classical segmentation by weighted aggregation algorithm and hierarchical graph cut algorithm.At last, we summarize the presented work and further discuss the future work on texture segmentation.