[Objective]Fuzzy C-means(FCM)performs the clustering directly on original data,and is sensibly influenced by noises and outliers.Currently for the purpose of tackling this issue,the most widely used method is based on mining results of data clustering structure and relaxing the fuzzy membership or local relationship of the sample points.Nevertheless,it can only reduce,but not completely eliminate,the effect of noises and outliers.To address this issue,herein we propose a novel clustering algorithm called truncated robust fuzzy C-means(TRFCM).[Methods]In TRFCM,the truncation technique is introduced based on the fuzzy local information C-mean(FLICM)model.The main idea of the proposed TRFCM is threefold:(1)by utilizing FLICM,the local neighborhood structure of sample points is preserved during the learning of data clustering structure;(2)on the basis of the clustering result of FLICM,original data is adjusted dynamically to meet the desired clustering structure;(3)an optimization framework is constructed to appropriately integrate(1)and(2).[Results]Proposed TRFCM is compared with algorithms developed in recent years.Our comparative experiments include categories,namely(1)parameter sensitivity and convergence analysis,(2)robustness,(3)image segmentation,(4)benchmark dataset and(5)computational time cost.For(1),within the appropriate range,TRFCM behaves insensitively to the parameter and can produce effective clustering results in most cases.Meanwhile,the clustering algorithm on each dataset can converge within 20 rounds of iterations.For(2),the accuracy of TRFCM reaches 81.55％,exceeding FLICM by 9.71 percentage points and clustering results approach closely to the real data distribution,demonstrating the robustness of TRFCM to noises.In(3),all compared algorithms do not behave sufficiently accurately to segmenting the image to a certain extent,and some of them endure troubles such as incomplete segmentation of environment,misclassification of different parts into the same class,and overlap among different classes.Instead,TRFCM avoids all these problems and produces satisfactory clustering results.To further compare the robustness of each algorithm to noises,we perform the image segmentation on images added with Gaussian noises with mean 0 and variance 0.05.Experimental results show that TRFCM performs optimally for the suppression of noise interferences.In(4),the clustering analysis is applied on Banknote Authentication,Wine,COIL20,WarpPIE10P,Yale and USPS datasets.During the experiment,10 repetitions of random initialization are conducted to measure the mean and standard deviation of three clustering metrics:ACC,NMI,and purity.According to experimental results,TRFCM achieves more satisfactory clustering results than other comparative algorithms do in those three aforementioned evaluation metrics.This outcome suggests that,in addition to the application of image segmentation,TRFCM also performs effectively for the division of real-world discrete datasets.In(5),TRFCM achieves more satisfactory clustering results compared to other algorithms under similar time costs.[Conclusions]A fuzzy clustering algorithm called TRFCM is proposed.Based on FLICM,the truncation technique is introduced to enable the improved TRFCM to dynamically adapt to original data and remove the interference of noises and outliers during clustering,so that more useful details for clustering are retained.TRFCM achieves meritorious results in the parameter sensitivity and convergence analysis,robustness testing,image segmentation experiments,benchmark dataset experiment as well as computational time-cost experiment,indicating the effectiveness of the algorithm.Inspired by these merits,previous classical fuzzy clustering models can be similarly modified by the truncation technique as an essential future research direction.
万方数据
维普
