He Research Of SAR Image Target Recognition Method Based On NMF

Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) as animportant way of SAR resources applications is playing an increasingly important rolein civilian and military fields. Feature extraction is the key technologies and coremission of the SAR ATR. Commonly used feature extraction methods from the field ofoptical image processing. Its need to be appropriately improvement before applied tothe SAR image. In this thesis, the special mechanism of the SAR imaging is used insparse NMF-based feature extraction method, combined with the smoothness constraintand semi-supervised constraints. Focus on mining the target feature informationcontained in the SAR image data. The specific contents are as follows:Introduced PCA and ICA feature extraction method which based on the nature ofthe matrix and used in MSTAR images. Followed by detailed non-negative matrixfactorization (NMF) method and derived convergence of NMF. Finally, all methodsapplied to the three types of tanks feature extraction.SAR image contains sparsity because of the special imaging environment andmechanism of SAR. The sparse information in an image can be extracted by improvethe NMF method. Two improved method based on the existing sparse NMF areproposed in this thesis, which are based on Smooth sparse NMF and based onsemi-supervised sparse NMF. Both improved method can effectively characterize thesparsity in the SAR image. The performance of improved method is superior to theexisting sparse NMF in convergence speed, features sparsity and features images.In order to verify the effectiveness of the improved feature extraction method, theParzen window characteristics Gaussian distribution function is used to calculate thecharacteristics pitch in this thesis:1. the change in distance between T72and BMP2at15degrees,2. the change in distance between BTR70in the class between15degreesand17degrees. Followed by the support vector machine (SVM) for classification andidentification of the target and calculate the classification and recognition rate at sametime.In contrast, property scattering center model-based feature extraction method is proposed in this thesis. The SAR image scattering centers can accurately extracted bythis method through set model, initial parameters of selection and parameteroptimization iteration. Compare the feature extraction results between semi-supervisedsparse NMF and property scattering center-based model extraction can be found that theextracted feature of the former can match the scattering centers preferably.