Improvent Of K-means Clustering Algorithm And Its Application

With the development of computer vision technologies, image segmentation has become a hotspot of image engineering. Image segmentation algorithms based on clustering analysis has already got great development since its birth. The K-means(KM) clustering algorithm is the most familiar and the most important algorithm. It has simple principle and is easy to implement. When the clusters is dense and have obvious difference, the KM clustering algorithm can achieve better result. However, the KM clustering algorithm has some weaknesses as follows:The number of clustering K and the initial centers must be determined before the algorithm is implemented according to the prior knowledge. The algorithm is sensitive to initial center and is prone to be trapped in the local optimum. The Moving K-means (MKM) clustering algorithm introduced fitness into KM clustering algorithm to improve the segmentation performance. The MKM algorithm was proven to be effective in avoiding the centre from being trapped in local minima. Furthermore, it requires much shorter processing time compared to the K-means clustering algorithm. But some members of centers with the largest fitness are enforced to be assigned as a member of a center with the smallest fitness. The enforcement process could reduce the accuracy of clustering. And the MKM algorithm is also sensitive to noise.This paper improved the KM and MKM algorithms based on deep study in them to enhance the accuracy of segmentation and improved the ability of resisting noise. The main contributions of the paper are as follows:Firstly, two modification versions are introduced:Adaptive Moving K-means (AMKM) clustering algorithm:One of the weaknesses of the MKM algorithm is that it obligates some members of the center with the largest fitness value to become the members of the center with the smallest fitness value. The process of reassigning these members in AMKM algorithm is different from the conventional MKM algorithm. It provides a solution to this problem by assigning these members of the center with the largest fitness to the nearest cluster depending on the minimum Euclidean distance. So it will significantly cut down the occurrence of classify noise data into inappropriate centers or clusters.Adaptive fuzzy moving K-means (AFMKM) clustering algorithm:The main idea is that it introduces the fuzzy concept into the MKM algorithm. It combines the properties of the fuzzy C-means(FCM) algorithm and the MKM algorithm and is less sensitive to noise.Secondly, two analyses named qualitative and quantitative analysis are applied to track the difference in behavior of our proposed clustering algorithms and several conventional clustering algorithms. We have tested the algorithms on standard images and noisy images. The comparison of subjective evaluation and the values of three unsupervised evaluation indexes such as F、F’、Q proved that the proposed algorithms presented better visual quality in accuracy and timeliness.Thirdly, the proposed algorithms and the FCM algorithm are developed for segmenting the clinker from the whole image. Furthermore, the paper classifies the sampled clinker images according to the features of clinker. The valuation indexes, segmenting time and the recognition rate prove that the proposed algorithms are better suitable to be implemented in visual inspection of complex industrial process.As shown in our exam results, the proposed algorithms exhibite better visual quality besides the speedup of segmentation, reduction of being trapped in local minima and anti-noise performance etc. So they could have some scientific significance and application prospect of expanding the vision application area.