Improvement And Acceleration Of Superpixel Segmentation Algorithm

With the rapid development of computer vision and multimedia technology,the size of digital image is increasing,the traditional pixel based image segmentation algorithm is more and more difficult to meet the real-time demand,and the superpixel segmentation algorithm to some extent solves this problem.Therefore,the superpixel algorithm has gradually become the most commonly used segmentation method in computer vision direction.The Simple Linear Iterative Clustering algorithm is one of the most popular superpixel segmentation algorithms,which is widely used in various production and life scenes.As preprocessing steps in image processing and computer vision applications,superpixel segmentation should meet the requirement of real time in real computing,but the SLIC algorithm still has great space to improve the segmentation speed when processing massive image data.Therefore,in order to solve this problem,this paper mainly carried out the following research:(1)Firstly,the process of SLIC algorithm is analyzed completely,and the efficiency of its segmentation is discussed.Then we propose a clustering method based on 4-Labeled neighbors pixels according to the clustering characteristics of the SLIC algorithm.This algorithm divides pixels into and two categories by the way of sampling,and the clustering which belongs to the pixels of part is constrained by the pixels of .The algorithm reduces the average calculation times of each pixel and improves the segmentation efficiency of the algorithm without reducing the algorithm segmentation effect.(2)On the basis of the results(1),the pixels of are subdivided,and a more optimized pixel classification and clustering method 4LCP-SLIC algorithm is proposed.This algorithm makes more pixels do less similarity calculation,the average calculation number of pixels in the algorithm is further reduced and the segmentation speed is further improved.(3)Two kinds of similarity measurement formulae are proposed for two-part pixels in the result(1)and result(2)methods.According to the similarity principle of neighborhood,a method of similarity measurement based on color distance is proposed,which reduces the computation of two dimensions and improves the clustering speed of partial pixels.For the partial pixels,the redundant term in the similarity measure formula is deduced by theoretical deduction,and a method of optimal cluster prediction is proposed,which improves the clustering speed of the partial pixels.The combination of two similarity measure formulae makes the segmentation speed of the algorithm in the result(1)and result(2)be improved once again,which basically meets the requirement of real-time processing.