Improved Fuzzy C-Means Algorithm And Its Application In Image Segmentation

As one of the research focus on digital image of engineering, machine vision and related fields, image segmentation has become a key technology from image processing to image analysis and understanding, and its effect will directly affect the completion of follow-up task. As the complexity of real images and fuzziness, fuzzy clustering method based on image segmentation is being widely concerned and applied.On base of reviewing and analyzing the related scientific literature of image segmentation algorithm, Mean Shift algorithm and fuzzy clustering algorithm, this paper deeply study typical methods of image segmentation and several types of improved fuzzy clustering method of image segmentation in recent years, and mainly discuss the shortcomings of fuzzy clustering method while segmenting image with noise and similar color area. The main research contents are summarized as follows:First, this paper introduces Mean Shift algorithm in the traditional FCM algorithm, and using the advantage of Mean Shift which can quickly find the peak point and the space neighbor information while clustering, and proposes an improved FCM algorithm -- Mean Shift based fuzzy C means clustering algorithm. The results show that the improved algorithm reduces the influence of noise.Second, although the Mean Shift based fuzzy C means algorithm is robust to noise, the algorithm is isotropic and its effect is not ideal while segmenting image with slender regions and color similar regions. In order to retain more image information and to further improve the algorithm's noise immunity, on base of the algorithm, this paper also propose another improved algorithm -- anisotropic Mean Shift based fuzzy C means algorithm. In the iteration process the bandwidth of kernel function is updated by sample points'information in real time, and the segmentation effect has a strong noise immunity, at the same time the convergence rate is improved.Finally, in order to get optimum results for image segmentation, this paper also uses validity function to determine the optimal number of clusters, thereby the error caused by improper selection of cluster numbers is reduced. Application of the improved algorithm and the related simulation experiments show that the validity of the proposed method.