| Image segmentation is an important pre-processing step for image recognition,image analysis and image understanding.Due to various subjective and objective factors,digital images usually have certain fuzziness.Among the many image segmentation algorithms based on uncertainty theory,the evidential c-means(ECM)clustering algorithm based on evidence theory describes the belief degree of pixels to each cluster through the basic belief assignment function.In the credal partition,pixels can not only belong to a singleton cluster or noise cluster,but also can belong to a meta-cluster united by several singleton clusters.This additional flexibility is conducive to more reasonable expression and processing of fuzzy information in the image.However,ECM does not consider the spatial information of pixels and cannot effectively deal with the noise points in the image.It is difficult to obtain satisfactory results by directly applying ECM to image segmentation.To address the above problem,three improved algorithms are proposed in this thesis,which are:(1)Aiming at the problem that ECM does not introduce the spatial information of pixels,resulting in the noise points cannot be effectively identified only by the global fixed noise distance and the pixels belonging to the meta-cluster and noise cluster in the clustering results cannot be accurately recovered,an evidence c-means clustering combining spatial information(ECM＿S)for noisy image segmentation algorithm is proposed firstly.ECM＿S utilizes the neighborhood information of pixels to construct an adaptive noise distance to improve the ability to identify noise points.Then,the original,local and non-local information of pixels are introduced into the objective function through adaptive weights to enhance the robustness to noise.Meanwhile,the entropy of pixel membership degree is used to design an adaptive parameter to solve the problem of distance parameter selection in credal c-means clustering(CCM).Finally,the Dempster’s rule of combination was improved by introducing spatial neighborhood information,which is used to assign the pixels belonging to the meta-cluster and the noise cluster in the credal partition to a specific singleton cluster to complete the image segmentation.Experimental results indicate that the ECM＿S has better applicability in noisy image segmentation.(2)Aiming at the problem that category information of neighboring pixels cannot be utilized to iteratively update the noise distance and recover the noise cluster in ECM＿S,a self-updating noise distance based adaptive kernel evidence c-means clustering combining spatial information(SNDAKECM＿S)for noisy image segmentation algorithm is proposed.In the iteration,the basic belief assignment function of the neighborhood pixels is utilized to construct the self-updating noise distance and the adaptive recovery factor,enabling SNDAKECM＿S can adaptively adjust the distance from the center pixel to the noise cluster and the weight of the recovery factor according to whether the neighborhood pixels contain enough category information to recover the center noise point.Concurrently,the sum of the maximum noise probabilities of all pixels is used for the adaptive selection of the scale parameter in the Gaussian kernel function.Experimental results indicate that SNDAKECM＿S can effectively overcome the impact of noise and uneven illumination on segmentation accuracy and the effectiveness of self-updating noise distance and adaptive recovery factor is verified.(3)Aiming at the problem that SNDAKECM＿S cannot be used for noisy color image segmentation,a mahalanobis distance based kernel suppressed evidence c-means clustering combining spatial information(MKSECM＿S)for noisy color image segmentation is proposed.MKSECM＿S introduces the color information of pixels in the initialization of noise distance,which enables the algorithm to effectively identify color noise points.In addition,the Mahalanobis distance is used to replace the Euclidean distance in the scale parameter adaptive Gaussian kernel function,which can make MKSECM＿S fully consider the correlation of the feature attributes of each color channel when measuring the difference between pixels and clustering centers.In order to reduce the number of iterations of MKSECM＿S,the concept of rejection interval in evidence theory is utilized to realize the adaptive selection of the suppressed factor.Experimental results indicate that MKSECM＿S has a better segmentation effect on noisy color images and the effectiveness of the adaptive suppressed factor is verified. |
