SAR Target Recognition Based On Classifier Fusion Methods

As a kind of active microwave sensor,Aperture Radar Synthetic(SAR)has the ability of all-time,all-weather and long distance observation.Its powerful penetration ability makes it able to obtain the information which can not be obtained in the general image.Through the interpretation of SAR images can get the information that can not get in other type images,which has an important military and civil value.As a part of SAR image interpretation,SAR target automatic recognition has been widely concerned by scholars at home and abroad.The traditional SAR ATR method uses a single classifier to identify the single feature,which has some one-sided and limitations,and often gets poor recognition results.In view of this problem,information fusion technology is used to integrate multiple types of classifiers with various features,which can achieve complementary of classifier performance,and improve the indentification accuracy and stability of the classifiers at the same time.The information fusion technology is applied to SAR ATR,through the extraction of targets' different types of features,constructing a plurality of different types of single classifier,and using automated parameter optimization method to get the best single classifier,combination classifer is constructed from these single classifiers selected.The output information of single classifier can be divided into three levels: the abstract level,the rank level and the measurement level,and the classifier fusion method is also used in these three level,which includes the class label methods of the abstract level,the class ranking methods of the rank level and the soft output methods of the measure level.In this paper,the samples used for experiment are from a TerraSAR scene image.By extracting the geometric feates,LBP texture features,PCA transform features and Hu invariant moments features,mulitple single classifier is constructed with the logic regression algorithm and support vector machine algorithm.The select the best single classifiers as the base classifier,using the weighted voting method,ranking based method and averaged Bayes fusion method to achieve the combinated classifier.In the experiment,a certain percentage of samples are randomly selected for training to get the classifier model.Then the whole samples are tested and identitied,which are repeated for ten times.The experiment results show that classifier fusion methods can overcome the inaccuracy of single feature classification model,which can not only improve the recognition accuracy of the classifier for each class samples,but also greatly improve the stability of the classifier,and has broad application prospects.