Image Segmentation Of Improved Spectral Clustering

Spectral clustering as a class of new and effective clustering method has been widely concerned and successfully applied in image segmentation. The image segmentation algorithm, based on initialization-independent multi-parameter kernel spectral clustering which was raised by Ma Xiuli, has been researched, improved and successfully applied in texture image and SAR image segmentation in this dissertation. The main achievement of this dissertation can be summarized as follows:1. Spectral clustering uses the image similarity matrix in image segmentation. How to construct a similarity matrix which expresses more information of image has a great influence on algorithm performance. Similarity matrix which was constructed by the mean of wavelet transform coefficients in original method, is constructed by the mean of dual-tree complex wavelet transform coefficients in this dissertation. Dual-Tree complex wavelet not only maintains good time-frequency localized analysis the capabilities of the traditional wavelet, but also has a good direction analysis ability. It can reflect that image changes on different resolutions along several directions, and describe the geometry of image better.Initialization-independent K-Harmonic means (KHM) was used to achieve better and stable clustering performance in original method. But it only deal with the low-dimensional data better than K means (KM). In this dissertation, the dimension of complex wavelet feature is higher than the wavelet's. Thus, KHM is not suitable. In order to relieve the problem that spectral clustering is sensitive to initialization, K-means based on optimization strategy has been adopt to realize the image segmentation based on complex wavelet feature and spectral clustering. A large number of simulation experiments in Brodatz library show that:In terms of fine texture, rough texture, non-uniform texture and blending textures, our method is better than the method in literature [57] according to the visual effects and evaluation indicators.2. When this method is used in SAR image segmentation, the edge positioning is inaccurate.Thus, we combine this method and watershed method, propose the strategy of feature fusion, and achieve the image segmentation based on threshold marker-based watershed and sepctral clustering. This method uses watershed for pre-segmentation, which can locate the edge accurately. By the pre-segmentation image, we can obtained the complex wavelet feature vector in each divided region to reduce the dimension of similarity matrix. The SAR image simulation experiments show that:the method in this dissertation is better than that in the literature [61] and [62].