SAR图像自动目标识别算法研究

发布时间：2018-06-24 02:39

本文选题：SAR图像 + 目标识别　；参考：《哈尔滨理工大学》2014年硕士论文

【摘要】：合成孔径雷达(Synthetic Aperture Radar, SAR)图像自动目标识别(AutomaticTarget Recognition, ATR)技术是在无人为干涉的情况下，，利用采集到的SAR图像数据，准确的找到感兴趣的区域，并提取目标的特征信息，识别出目标的类别属性。在SARATR体系中包含最主要的三个步骤，即SAR图像预处理，SAR图像目标特征提取和SAR图像目标分类识别。本论文针对这三个步骤，完成了以下任务： 1. SAR图像的预处理方法：本论文研究了一种基于复Contourlet变换和隐马尔科夫树(HMT)模型的SAR图像去除相干斑噪声方法。该方法利用复Contourlet变换的多尺度、多方向性和平移不变性的特点，将其与HMT模型相结合，从而能够准确地描述复Contourlet变换域系数在相邻尺度间的相关性。对实验结果的定量分析可知，该方法取得了良好的去噪效果。 2. SAR图像的特征提取方法：流形学习方法是模式识别的基本方法，本文应用的最大异类距离嵌入特征提取(Maximum Interclass Distance Embedding, MIDE)方法的创新性体现在既结合了主成份分析(Personal Computer Assistant, PCA)的差异特性和局部保持映射(Local Preserving Projection, LPP)的邻域信息，并同时融入了数据集的类别信息。利用此方法旨在找寻一个线性嵌入的映射，能够将投影后获得的不同类别的SAR图像特征彼此互相远离。 3. SAR图像目标识别算法：文中提出了一种基于神经网络集成模型的SAR图像分类识别算法。该方法很好的克服了SAR图像对方位角敏感的问题，针对同向目标的特征空间训练一个神经网络实现目标分类，并使用另一个二级神经网络对多个单向目标识别器的识别结果进行结合，提高识别精度。将SAR图像自动目标识别三个阶段的三种方法同时应用在SAR图像自动目标识别中，通过仿真实验，定量分析可知，其最终的目标识别率高于其它较为优秀的算法所达到的目标识别率，证明该论文提出的算法是切实可行的。
[Abstract]:The automatic Target recognition (ATR) technology of synthetic Aperture radar (SAR) image is to accurately locate the region of interest and extract the feature information of the target by using the collected SAR image data under the condition of no one interference. Identifies the target's category attributes. There are three main steps in SARATR system, that is, SAR image preprocessing, SAR image feature extraction and SAR image target classification and recognition. According to these three steps, this thesis has completed the following tasks: 1. The method of SAR image preprocessing: in this paper, a speckle noise removal method for SAR image based on complex Contourlet transform and Hidden Markov Tree (HMT) model is studied. This method combines the multi-scale, multi-directivity and translation invariance of complex Contourlet transform with HMT model, so it can accurately describe the correlation between adjacent scales of complex Contourlet transform domain coefficients. The quantitative analysis of the experimental results shows that the method has achieved a good denoising effect. 2. The feature extraction method of SAR image: manifold learning method is the basic method of pattern recognition. The innovation of Maximum Interclass distance embedding (MIDE) method used in this paper lies in the combination of the difference characteristics of personal computer Assistance (PCA) and the neighborhood information of Local preserving Project (LPP). At the same time, it incorporates the category information of the data set. Using this method, we can find a linear embedded map, which can separate different kinds of SAR image features from each other. Target recognition algorithm for SAR images: a classification and recognition algorithm for SAR images based on neural network ensemble model is proposed in this paper. This method can overcome the problem that SAR image is sensitive to azimuth, and train a neural network to classify the target in the feature space of the same target. Another two-level neural network is used to combine the recognition results of multiple unidirectional target recognizers to improve the recognition accuracy. The three methods of automatic target recognition in SAR image are applied to the automatic target recognition of SAR image at the same time. The final target recognition rate is higher than that achieved by other better algorithms, which proves that the algorithm proposed in this paper is feasible.
【学位授予单位】：哈尔滨理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【参考文献】