Parallelization Research Of Density-Based Clustering Algorithm And Its Application In Retinal Blood Vessels Extraction

As a kind of data statistical analysis method, Clustering is widely used in Machine Learning, Image Processing, Complex Network Analysis. Due to the differences in the clustering process, clustering algorithms can be classified into partition-based, hierarchy-based, density-based, grid-based and model-based method. Density-based clustering algorithm is favored by researchers, because it is not sensitive to noise and can be used to detect the clusters with arbitrary shape. However, when dealing with large amount of data or multi-dimensional data, density-based clustering algorithm has the problem of high overhead in distance-calculation. Algorithm parallelization is an effective way to improve computing performance. And the development of GPU general-purpose computing technology and CUDA support the research of algorithm parallelization.This paper focuses on the study of Density Peaks clustering algorithm. The improved Density Peaks clustering algorithm is provided and its parallelization design based on GPU is implemented. Moreover, combining the application in Image Processing, this paper designs Density Peaks clustering parallel retinal blood vessels segmentation algorithm.Firstly, the paper improves the Density Peaks clustering algorithm by adjusting the density-calculation method and the distance-definition, taking Gaussian Kernel and KNN to compute the local density of data points, introducing the descending order of density to define distance, for decreasing clustering mistakes and improving the accuracy of clustering. Secondly, the paper puts much emphases on researching the improved Density Peaks clustering algorithm’s parallelization design and its realization on GPU under the CUDA architecture. By taking advantage of potential parallelizable characteristics in algorithm, the same operation for each data point, the large number of computational tasks, such as calculating distance, density and decision value, can be transferred to GPU platform to parallel processing for enhancing the efficiency of the program. Finally, Density Peaks clustering parallel algorithm is applied to retinal blood vessels segmentation. The pixel level hybrid features library is constructed by extracting SWT, WLD, Gabor response information of pixels in the image. Density Peaks clustering parallel retinal blood vessels segmentation algorithm clusters the pixels in the multi-dimensional feature space to realize the segmentation of blood vessels.The paper has made the contrast experiments among Density Peaks clustering algorithm, improved Density Peaks clustering algorithm and Density Peaks clustering parallel algorithm, and the application-experiment of retinal blood vessels segmentation. Experimental results show that the Density Peaks clustering parallel algorithm can effectively improve the speedup.