Image Segmentation Algorithm Based On Fuzzy Clustering Analysis

Fuzzy clustering analysis has the ability of classifying samples with uncertainty and has been widely applied in image segmentation in recent years. An improved fuzzy clustering algorithm was proposed in this paper based on classical fuzzy c-means (FCM) algorithm and the validity of the new algorithm was verified.Firstly, an open problem of standard FCM is to determine the number of clusters in advance. Different images have different amounts of cluster, thus selecting appropriate number of clusters is crucial to the results of image segmentation. Usually, researchers determine the number of clusters by the number of histogram peaks of an image. However, influenced by noise, this method easily yields error because the concept that a pixel belongs to a gray level is not clear. Considering that fuzzy set theory can reflect the degree of a pixel belonging to a gray level, fuzzy histogram and fuzzy histon histogram are constructed respectively in the gray images and color images by replacing impulse function to Gaussian function and making full use of gray information and spatial information of pixels.Secondly, although gray information and local spatial information of each pixel can restrain the influence of noise in some way. Using standard FCM algorithm with local spatial information of pixels cannot obtain significantly better performances of segmentation when the images are badly disturbed by noise. To solve this problem, a non-local spatial adaptive FCM algorithm based on fuzzy histogram was presented by embedding the non-local spatial information of the pixels into the objective function. We derived new iterative formulas of both membership matrix and cluster center from the improved objective function. Experimental results show that the proposed algorithm is better than standard FCM algorithm in the quality of segmentation.Finally, an adaptive FCM algorithm was considered in Lab color space, in which all possible homogeneous regions labeled by cluster centers can be found by using fuzzy histon threshold technology. The proposed adaptive FCM algorithm improves the compactness of all homogeneous regions and successfully achieves image segmentation. Moreover, the experimental images of my proposed algorithms come from Berkeley image database, and my algorithms obtain an ideal experimental result. Considering that color is similar but semantic is irrelevant in complex natural images, we add textural feature into the feature set of pixels which can effectively distinguish complex images. Thus, a new FCM segmenting algorithm suitable for color images is put forward combining the features including lightness, color and texture. Compared with standard FCM algorithm, the new algorithms perform better in both segmentation quality and anti-noise ability for color images.