SAR Image Target Recognition Research Based On The Projection Features

Synthetic aperture radar (SAR) with high resolution, strong penetrating, long-rangeimaging and all-weather working abilities, it is more and more widely used in the fieldof military and civilian applications. In recent years, the SAR Automatic TargetRecognition (ATR) technology has become the world’s hot issue. This thesis researchesthe SAR image pre-processing and feature extraction method. The main contents are asfollows:1. It researches SAR image pre-filtering algorithm and two-parameter CFARsegmentation algorithm. This thesis proposes the "islands" and "holes" processingmethod which based on regional labeling algorithm. Compared with the traditionalmethod, it is not only preserving image details, but also restoring the lost informationduring image filtering and segmentation. Furthermore, it uses unified image resolution,normalized energy and enhanced power transformation based on image centroid, so thatit improves the probability of the target image. By experiments, we finds out the optimalpower transform coefficients.2. Because of the shortage in traditional PCA and LDA feature extraction methodswhich are based on the one-dimensional vectors, this thesis studies the methods of2DPCA and2DLDA based on two-dimensional matrix. Experiment results show thatthe2DLDA has optimal performance compared with that of2DPCA by atwo-directional compression. By studying the sample dependence and orientationsensitivity, it proves that the feature extraction method based on two-dimensional matrixis better than the one based on one-dimensional vectors.3. It proposes a new feature extraction method by combining2DPCA with2DLDA. This method uses2DPCA to decrease the dimensionality and2DLDA toincrease sample divisibility ability. So that it combines the advantages i.e.dimensionality reduction and data structure as well as reserving the superiordiscrimination ability in2DLDA method, and finally more information in the targetfeature can be obtained by this algorithm.The experiment also verifies that thecombination method recognition rate are higher than that of a single feature extractionmethod.