Research On Color Image Segmentation Based On Pixels Clustering

With the continuous development and mature of digital imaging technology and image equipment, the related processing technologies of nature images have become auxiliary means of human intelligent activities. These technologies of image processing are no longer stalled on the level of storage, transmission and exchange, but developing to diversification and intelligent. The final target of image engineering is to discover and abstract high-level knowledge and develop cognitive competence from original images. As a key step towards this goal, image segmentation is the main research content of image processing and low level computer vision tasks. Because images are diverse and complex,image segmentation is a issue full of uncertainty as the premise of image understanding. Image segmentation involves mathematics, physics, statistics, psychology, cognitive science, signal analysis, machine learning theory, and no segmentation method is perfect and universal. That is why it is highly concerned by the researchers. In recent years, with the development and mature of the Internet, mobile devices, and data compression technologies, natural color image segmentation has gradually become the focus issue of segmentation. Automatic segmentation of nature color image is the basis for natural image understanding, retrieval, and annotation. It can achieve image compression and save the storage space, and it is widely used in industrial inspection, military reconnaissance, biomedical, meteorological traffic, scientific visualization etc.With analysis of color image segmentation task, this dissertation firstly summarize the essence of segmentation task as pixels clustering of image plane, and give analysis of the relationship between image segmentation and pixel clustering technology from the two aspects. The emphasis work of this dissertation is clustering segmentation technology, the self organizing map network, regional growing method, pixels clustering on major colors, Mean-Shift and texture segmentation model have been studied and explored by analyzing the distribution characteristics of the color image and combining with the statistical features. All these clustering segmentation methods are integrated with distribution features of pixels, the image blocking strategy, visual saliency theory and texture computing. The major results and contributions in this dissertation are as follows:1、Self organizing Map(SOM) network has natural ability of data classification, and how to determine the structure and scale of the network is always a difficulty problem of clustering segmentation. To solve the problem, we present a scale estimation method of SOM network based on the computing of pixels distribution in the feature space. The algorithm firstly compute pixel distribution statistics in HSV space and construct the network by estimating the scale of the SOM network based on maximums in color histograms, then it trains the neural network to a stable network based on sampling pixels from average grids of image. Each pixel is classified according to the approximate degree between image pixels and the neuron connection weight vector. Finally, we use the spatial consistency to adjust pixels clustering as segmentation results. The comparisons of visual and quantitative experiments with other kindered algorithms prove that the proposed method can achieve excellent segmentation quality.2、Region growing segmentation algorithm has advantages in maintaining regional connectivity, but the seeds selection and growth sequence restrict the universality of region growing algorithm. In this paper, we propose a new type of segmentation algorithm of automatic seed pixels selection and parallel growing strategy. First, we divide the image into blocks, and select representative pixels in local grids as seed pixels in accordance with the maximum distribution of colors. Second, we classify all seed pixels into clusters by dynamic partition-based clustering and conduct growing for all clusters seeds to corresponding regions. Finally we merge small regions to final segmentation. This method integrates image blocking and clustering seeds pixels into a novel efficient region growing segmentation. The experiments show that the proposed method is an effective and stable algorithm.3、Accurate classification of pixels is affected by many factors and it has not been solved perfectly. As a measurement of salient degree of pixels in the image, the visual saliency feature is positive for classification of pixels. By analyzing of various types of salience calculation methods, we propose a novel segmentation strategy based on salience fusion and define two kinds of suitable saliency map for pixel classification. The salient maps are combined with Mean Shift and major colors clustering segmentation respectively and two color image segmentation methods are obtained by fusing salient features. In the first algorithm, we compute salient map by fusing regional contrast of quantized image and pixels-region contrast, next the fused saliency map is integrated into Mean-Shift segmentation algorithm as characteristic vector. Another algorithm computed salience map by fusing region contrast of Mean Shift segmentation results and global contrast of pixels, then the saliency map is introduced to major color density clustering algorithm as the regulate factor to adjust the clustering results on the basis of the differences of pixels in the salience map. The experiments and comparison confirm that the salience map computing methods in this paper can improve the classification results of pixels effectively, and both methods obtain good segmentation results.4、In order to deal with the natural texture regions in the image, a texture segmentation method is proposed based on superpixel merging texture. Superpixels segmentation is gathering homogeneous pixels into regions according to certain rules ignoring whole image and it has been used in the preprocessing of many computer vision tasks. We firstly propose a texture feature extraction method based on Gabor filters, and the texture feature is integrated into the computing process of superpixels, which aims to protect the real boundary and texture integrity. Then, a color texture histogram feature is defined according to the main colors and their distribution characteristics in superpixels and the adjacent and similar superpixels are merged to get the final segmentation result by the similarity between histograms distance of regions using standard deviation constraints of all regions histogram features. The experiments and comparisions of our algorithm and related algorithms are conducted on test database, and the experiments results show that segmentation method has significant advantage on a subjective visual consistency and objective accuracy evaluation.