合成孔径雷达(SAR)是一种高空视觉系统，具有全天候、远距离、极强的穿透力和高分辨率等特点。针对SAR图像的目标识别已成为国内外研究的热点，而如何精确地提取图像特征和采用有效的分类识别方法一直是研究的重点。由于SAR图像与常见光学图像的不同，而且有严重的噪声，传统的特征提取方法应用于SAR图像有一定的局限性。支持向量机(SVM)是近些年发展起来的一种新的机器学习方法，它以统计学习理论为基础，能较好地解决小样本学习问题，由于其出色的学习和推广能力，支持向量机已经被应用到许多方面。 本文基于支持向量机方法对SAR图像特征提取和分类识别做了较深入研究。首先综述了SAR图像目标识别和支持向量机的研究现状和发展趋势；其次研究了SAR图像的特征，总结了SAR图像目标识别的一般流程；然后采用自适应滤波和马尔可夫随机场对SAR图像进行去噪和分割；在完成SAR图像的预处理后，进行了Hu不变矩稳定性实验，删除稳定性较差的矩分量。本文针对SAR图像目标的特征，结合Hu不变矩和三个仿射不变矩，重新组合了一组不变矩。利用重新组合后的不变矩对图像进行特征提取，该方法不仅具有旋转、平移和尺度不变性，还具有仿射不变性；最后将提取的特征值输入支持向量机进行训练识别，并通过有向无环图将两分类推广到多分类识别。 通过一系列实验表明，采用重新组合后的不变矩特征提取和支持向量机相结合的方法获得了较高的识别率，是一种有效的SAR图像目标识别方法，而实验中先对待识别目标进行相似度测试，降低了错误识别的概率，提高了系统的稳健性。

Synthetic aperture radar (SAR) is a system of high-altitude visual which has characters as all-weather, long distance, strong penetrable ability and high resolution. SAR target recognition has become a hotspot, the most important section is how to extract feature accurately and adopt effective classification recognition method. As the SAR image which is different from common optical image has serious problem with noise. And traditional image feature extraction method for SAR images is limited. Support Vector Machine (SVM) based on the statistical learning theory is a new method of machine learning. SVM which can settle small example problem well has been applied to many fields for its excellent learning and generalizing ability. This thesis mainly studies feature extraction and classification recognition of SAR image based on SVM. Firstly, reviews the research status and development trend of SAR target recognition and SVM. Secondly, studies the feature of SAR image and summarizes the general process of SAR target recognition. Then, we carried on the image de-noising and segmentation using adaptive filter and Markov random field. And the Hu invariant moments stability experiment is carried out, We delete the components of poor stability. According to the feature of SAR target, we recombine a set of invariant moments by combining Hu invariant moments with affine invariant moments to carry out feature extraction. This method not only has the characteristics of scale, translation and rotating invariance, but also affine invariance. Finally, introduce the SVM to training and recognition, then extend the binary classification to the multi-classification by directed acyclic graph. In series of experiments, the excellent recognition rate is achieved by using the recombined invariant moments feature extraction, therefore this method is valuable. We test the similarity for being identified target before recognition for reducing the probability of misrecognition and improving the robustness of the system.