SAR Image Classification Based On Submodular Dictionary Learning

Synthetic Aperture Radar(SAR)is a high-resolution radar system.Its imaging is basically not affected by weather,light and other factors.It can monitor targets all day and is widely used in agriculture,military,geological detection and other fields.As an important step in image interpretation,SAR image classification can provide the overall structure information and highlight the regions of interest,which plays an important role in subsequent image interpretation.In recent years,dictionary learning and sparse coding are more and more widely used in image classification.This paper focuses on the task of classification of SAR images,a novel submodular dictionary learning method is proposed,which helps to efficiently learn a compact and discriminative dictionary for sparse coding.Based on the study of the traditional image classification methods,this paper uses submodular theory to improve some of the methods:First,an improved method for calculating spatial pyramid features based on sparse coding is proposed.The origin method of calculating the dictionary is in an iterative way,the more iterations,the longer time of dictionary learning.In this paper,we use the submodular clustering to calculate the dictionary,which greatly reduces the time of learning a dictionary.Second,a novel submodular objective function is proposed for clustering on a graph.The function contains three terms:entropy rate of random walk on a graph,a discriminative term and a balancing term.The entropy rate term encourage homogeneous and compact clusters,the discriminative term makes each cluster has higher class purity,and the balancing term makes clusters have similar size.Third,we propose a new way to compute dictionary items.Using the expectation of the elements with the largest number of target classes in the cluster members instead of all the members,which maximize the using of main information of clusters,so that a more discriminative dictionary can be achieved.This paper combines the theory of graph theory,submodularity,dictionary learning,sparse coding,etc.Firstly,extract Scale Invariant Feature Transform(SIFT)features of image and obtain their spatial pyramid representations,which are used to construct a graph model.Then,by maximizing a submodular objective function,clustering the graph model,a dictionary is obtained.Finally,using the Support Vector Machine(SVM)and sparse coding to achieve SAR image classification.Compared with the traditional methods,the classification result achieved by our method has a higher classification accuracy,and the consistency of homogenous regions is better.