SAR Image Recognition Based On Subspace

As an all-time coherent imaging system, Synthetic Aperture Radar(SAR) is an important method of acquiring remote sensing information. SAR has been widely used in military and civilian field. Automatic Target Recognition(ATR) technology of SAR needs no human interference, which means computers alone can complete classification and recognization of SAR imags automatically. Consequently, more and more countries have paid their attention to the research of SAR ATR. This thesis focuses on subspace-based feature extraction methods of SAR images.The main contents are as follows:1. The SAR imaging is quite different from some other imaging techniques, and the original images of that are always mixed with a lot of correlated speckle noise, shadows, targets and backgrounds. So it is necessary to preprocess the images before feature detection. In order to make the target more identified, firstly changing the multiplicative noise into additive noise by logarithmic transformation. Secondly, selecting the filter based on the wavelet transform to filter noise, because experimental results show that this wave filter can suppress speckle noises as much as possible and the edge details can be retained at the same time. And then, compressing the contrast range of image to improve the identification ability of target by power transformation. After that, Segmenting image by using two-parameter constant false alarm rate(CFAR). What’s more, Resolution unification and energy normalization are all taken in the subsequent operations.2. When been used to solve multiclass classification problem, Linear Discriminant Analysis(LDA) and Two-Dimensional Linear Discriminant Analysis(2DLDA) will cause sub optimality problem which can adversely affect the recognition performance, so we studied the weighted 2DLDA. The weighted 2DLDA gives smaller weights to those edge classes and wild points which are unconducive to classification. Then the projection obtained by this way will be easier to distinguish the categories that were not well differentiated before. Therefore the final recognition performance is better than 2DLDA. Based on this, a new method that combines 2DPCA and weighted 2DLDA is proposed. This new method first use 2DPCA to reduce dimensions, then use weighted 2DLDA to decrease the discrepancy within classes and increase the discrepancy between classes. It has a better performance than other methods by complementing each other’s advantages.3. This thesis proposes a new feature extraction method that combing 2DPCA and 2DLPP together. This proposed method takes both advantages of 2DPCA and 2DLPP. Firstly, this method makes use of 2DPCA to acquire the global structure information of the target, and then it maintains the local structure by 2DLPP. Performance of targets identification would be improved by the procedures above.