Improving Spectral Clustering Algorithm Based SAR Spill Oil Image Segmentation

Abstract

The classic spectral clustering algorithm has a superior performance in the category in any shape of data collection, but the computational complexity of the classic spectral clustering algorithm is very high. In the case of limited computer memory and computing speed, we only can process low-dimensional image. The Nystrom algorithm uses sampling estimate the global sample in order to solve the problem of high computational complexity. However, the synthetic aperture radar (SAR) oil spill images have the large amount of data and the oil targets often present the case of non-uniform gray. Simple to use Nystrom algorithm will reduce the segmentation accuracy. To solve these problems, this paper adopts a grid method that the original image was divided into several sub-images first and then processed separately. This paper also design an algorithm to ensure that labeling of the classification labels for each sub-images are identical in the end. At the same time, the grid method also can reduce the computational complexity and the calculation amount of data storage. In each sub-image, random sampling method will further reduce the computational complexity. Instability effects cased by random sampling will be solved by using multiple sampling methods combined with the majority vote. Simulation results show that the improved spectral clustering algorithm combined with the grid method can reduce the computational complexity and improve the segmentation accuracy significantly.