|
Clustering technology as one of the automatic access to information technology in the field of speech recognition, image processing, has been widely used. Compared to the commonly used clustering algorithms, spectral clustering has to be in any shape of sample space clustering, and the ability to converge to the global optimal solution such advantages. Thus in recent years, has been a hot international machine learning research in this field. Image segmentation, as the cluster analysis is an important application field of research, by the image processing to the image analysis process is a critical step, occupies a very important position in the image project. Currently, medical, military, weather, traffic, and many social production and living a good application. The spectral clustering data sample a wide range of adaptability advantages, cluster analysis is a very viable image segmentation ideas based on the pixel and its characteristics. This paper attempts to spectral clustering algorithm for image segmentation, and on some issues, research and analysis. This study is based on spectral clustering method of image segmentation, improved on spectral clustering algorithm to be applied to the experimental analysis of image segmentation, especially in texture image segmentation, texture, color, and spatial characteristics, integration Finally, the experimental results are analyzed and compared. The research can be divided into the following aspects: (1) describes the basic theory of spectral clustering algorithm, the current research situation; spectral clustering algorithm can not be applied to large-scale data processing limitations using sampling based Nystrom algorithm method clustering algorithm combined with traditional spectral clustering in time and space complexity compared; adaptive parameter algorithm Gaussian kernel parameter sensitivity, a combination of neighborhood information Select a method to avoid manually adjust, experiments show that this adaptive spectral clustering algorithm based on the Nystrom sampling methods achieve good segmentation effect in the image segmentation experiments. (2) sets out the basic theory of semi-supervised learning, including commonly used several semi-supervised algorithm, combined with the characteristics of spectral clustering, spectral clustering algorithm is extended into a form of semi-supervised rely add pairs in the clustering limit as priori information to improve performance of clustering algorithms. The combined adaptive spectral clustering algorithm theory based on the Nystrom sampling method, forming a semi-supervised adaptive algorithm, the experimental results show that, add a semi-supervised image segmentation results significantly better than ordinary spectral clustering method. (3) characteristics for texture image is more complex, containing color, texture, and spatial characteristics, high dimensionality of the feature space, try using spectral clustering method to texture image segmentation, which use Gabor filters to extract texture characteristics. Design a group contains four scales, the filter group of the six directions, extraction multidimensional texture features, and then smoothing filter to obtain a more stable texture features proposed for the redundant information contained in the feature space using PCA method extracts the primary ingredients, and finally, according to the texture features and spectral clustering algorithm for image segmentation. In the experiment, respectively, using the spectral clustering algorithm for image segmentation based on texture features and color features Finally, after the integration of the texture, color and spatial characteristics of image segmentation.
|