Research On Military Target Recognition Method Of SAR Images

As an active coherent imaging radar, Synthetic Aperture Radar (SAR) is a very important means for remote sensing information, which is capable of imaging a large area of earth all-time and all-weather. It’s widely applied in the field of civil and military. The most important application of SAR in the military field is achieving detection and identification of the specific military target. So SAR image target recognition has been widespread concerned, and thus became a hotspot to domestic and foreign scholars. Around the military vehicles SAR image target recognition, this thesis mainly research the SAR image preprocessing and feature extraction, select six kinds of vehicles SAR image data from MSTAR public data source to complete target recognition experiment, test the effectiveness of pretreatment methods and analyze the influence to identification results from extraction methods. The main contents of this thesis include:For requirements of removing SAR image background interference and enhancing the classified information, the systemic pre-processing for MSTAR data is adopted. First, the SAR image pre-processing uses Lee filter and logarithmic transform to reduce noise; then the images are processed by threshold segmentation, morphological filtering, and geometric clustering to extract information of the target area, so background interference is eliminated; the target mask images are processed by power transformation, energy normalization and two-dimensional Fourier transformation, and thus the images are more conducive to feature extraction and classification.According to the theory of pattern recognition, SAR image target recognition should be one of the supervised machine learning The thesis selects six kinds of supervised learning feature extraction methods for the experiment of SAR military vehicle recognition, divides them into the methods based on image vector model (including PCA, LDA, LPP, NPE) and the methods based on image matrix model (including2DPCA,2DLDA), combines these methods with nearest neighbor classifier for the experiments of the preprocessed SAR image, and then analyze the applicability of various methods to SAR image target recognition by the experimental results.The characteristics of various classic feature extraction methods on SAR image target recognition are summarized. On this basis, a two-stage two-dimensional locality discriminant embedding feature extraction method is proposed, and combined with new methods of pretreatment for experiments. The samples of SAR images were processed as the form of matrix directly and the features were extracted by two stage method, where the image matrix was projection transformed from two directions of row and column respectively. Thus, the curse of dimensionality and the small sample size problem were avoided, which were always appeared in LDE method where the samples were transformed to vector. Meanwhile, the features could be more discriminatory. After feature extraction, the nearest neighbor classifier was applied for target classification, and the results show that the method is effective and advantageous.