The Key Technology Of Automatic Target Recognition In High Resolution SAR Images

Synthetic aperture radar(SAR) has been widely used in civil and military fields for the reason that its imaging is not affected by weather, illumination etc. With the further development of SAR technology, SAR image based on the traditional manual interpretation gradually transformed into SAR image automatic target recognition(SAR ATR), solving the problem of heavy workload and subjective mistakes in understanding brought by artificial interpretation when the identifying target has a lot of data and rich details. This thesis mainly studies the key technologies of SAR ATR, including three parts: filtering of SAR image, feature extraction of SAR image and classifier design for SAR image. A detailed analysis is given on the algorithms involved in each part, respectively from theory and experimental data.An anisotropic diffusion denoising algorithm based on difference curvature for SAR images is presented. The imaging characteristics of SAR determine that the SAR image is affected by speckle noise and is in accordance with the multiplicative model. In the process, through a logarithmic transformation, the multiplicative model is transformed into additive model obeying Gauss distribution. Studies the denoising algorithm based on partial differential equation, according to fact that the traditional P-M, mean curvature, Gaussian curvature denoising algorithms cannot retain information of image edges and details, puts forward the improvement of anisotropy diffusion method based on differential curvature driven. The method can distinguish relatively well the edge, the isolated noise and flat areas of the image, therefore, can retain the image edge while denoising.A feature extraction method based on K-L transform for SAR images is researched. Recognition efficiency is extremely low when the SAR image classification is carried out directly after denoising. Therefore, the image feature extraction can be provided as necessary before classification, in this way, representative feature is extracted and further denoising is achieved. Studies and analyzes the PCA algorithm. The algorithm using the K-L transformation principle, transforms the image into a one dimensional vector, and selects representative feature vector according to the size of the eigenvalues of the covariance matrix. Further studies the 2DPCA algorithm which is similar to the PCA algorithm. This algorithm directly uses the original image to do feature extraction, solving the problem that large amount of calculation is needed in large images processing by PCA.A composite SAR targets classification algorithm based on K-NN and SVM is proposed. Classifies targets using the projection data obtained from the feature extraction process, then mainly analyses three design methods of classifiers: the nearest neighbor algorithm, support vector machine algorithm and unequal interval optimal hyperplane of support vector machine algorithm. As the research targets are of many categories, one-to-one classification design is adopted, and combines each two of the categories of the training samples to respectively design two-class classifier, then votes from the results of each classifier.Combined with the characteristics of nearest neighbor and support vector machine, this thesis proposes an improved classifier design method. This method uses different classifiers on the basis of the different distribution of training data, which, on the one hand solves the problem that calculation and storage space is of large amount when a large number of sample data directly uses nearest neighbor method to do classification, on the other hand, avoids the problem that noise sample points near classification surface are easily classified in error using support vector machine method.