基于语义和脊波反卷积网络的SAR图像分割

发布时间：2018-04-13 01:37

本文选题：SAR图像 + SAR图像分割　；参考：《西安电子科技大学》2016年硕士论文

【摘要】：SAR图像分割是SAR图像处理的基础,分割质量的好坏直接影响后续的分析、识别等。传统的SAR图像分割方法主要分为基于特征的方法和基于统计模型的方法。基于特征的方法中,特征的好坏往往成为整个SAR图像处理系统性能的瓶颈。传统的基于特征的方法需要靠人工经验来提取样本特征,但是人工选择特征是一项复杂而又难以控制的工程。因此,用于自动提取SAR图像本质特征的方法亟待出现。深度学习是一种能够自动提取图像特征的方法,但是传统的深度学习模型对SAR图像这种结构复杂的图像处理效果并不理想。本论文针对上述难题提出了一种基于语义和脊波反卷积网络的SAR图像分割方法,创新点如下:(1)在脊波冗余字典中随机选取一组基元对反卷积网络每一层的滤波器进行初始化,得到脊波反卷积网络。脊波冗余字典具有丰富的尺度、方向和位置信息,能够更好的匹配图像,捕捉更多的图像结构特征,因此相比于用随机初始化方法以及用高斯初始化方法对滤波器进行初始化,用脊波冗余字典初始化滤波器可以学习到更好的SAR图像特征。基于语义空间将SAR图像划分为聚集区域、匀质区域和结构区域。对空间上不连通的SAR图像的聚集区域或匀质区域分别训练一个脊波反卷积网络,通过脊波反卷积网络的特征学习,得到代表各区域的特征表示。然而空间上不连通的聚集区域或匀质区域可能是同一类地物,这就需要根据区域特征的相似性来对聚集区域或匀质区域分别进行区域合并,得到SAR图像聚集区域和匀质区域的分割结果。(2)与聚集区域结构特点不同的是,匀质区域具有微纹理特点,区域内明暗变化不强烈,使用基于脊波反卷积网络的方法对其进行分割得到的分割效果并不理想,而灰度共生矩阵对处理纹理结构具有很大的优势。因此,本文使用基于灰度共生矩阵和层次聚类的方法对匀质区域进行了分割:首先提取匀质区域的样本,获得匀质区域样本灰度共生矩阵的熵和相关性等二次统计量,计算匀质区域样本的灰度均值和均方差,将灰度共生矩阵的二次统计量和灰度均值均方差组成一个特征向量,对得到的特征向量进行层次聚类,得到匀质区域分割结果,取得了良好的分割效果。结构区域的地物一般为边界、线目标和独立目标,因此,对结构区域使用基于分水岭的方法进行分割,并整合聚集区域、匀质区域和结构区域的分割结果,得到最终SAR图像的分割结果。
[Abstract]:SAR image segmentation is the basis of SAR image processing, and the quality of segmentation directly affects the subsequent analysis, recognition and so on.Traditional SAR image segmentation methods are mainly divided into feature-based and statistical model-based methods.In the feature-based approach, the performance of the whole SAR image processing system is often the bottleneck.Traditional feature-based methods require artificial experience to extract sample features, but artificial selection of features is a complex and difficult to control project.Therefore, the method for automatic extraction of essential features of SAR images is urgently needed.Depth learning is a method that can extract image features automatically, but the traditional depth learning model is not ideal for the complex image processing of SAR image.In this paper, a SAR image segmentation method based on semantic and ridge deconvolution networks is proposed. The innovations are as follows: 1) A random set of primitives are selected in the ridgelet redundancy dictionary to initialize the filters in each layer of the deconvolution network.The ridgelet deconvolution network is obtained.The ridgelet redundancy dictionary has abundant scale, direction and position information, so it can better match images and capture more image structural features, so compared with random initialization method and Gao Si initialization method, the filter can be initialized.We can learn better SAR image features by initializing filters with ridgelet redundancy dictionary.Based on semantic space, the SAR image is divided into three regions: aggregation region, homogeneous region and structure region.A ridgelet deconvolution network is trained for the spatial disconnected SAR images in the clustering region or homogeneous region. The feature representation representing each region is obtained by learning the features of the ridgelet deconvolution network.However, the spatially disconnected agglomeration or homogeneous region may be the same type of feature, so it is necessary to merge the agglomeration region or homogeneous region according to the similarity of regional characteristics.Different from the structure of the clustering region, the homogeneous region has the characteristics of micro-texture, and the change of light and shade in the region is not strong.The segmentation effect based on ridgelet deconvolution network is not satisfactory, but gray level co-occurrence matrix has a great advantage in processing texture structure.Therefore, in this paper, the homogeneous region is segmented based on gray level co-occurrence matrix and hierarchical clustering. Firstly, the sample of homogeneous region is extracted, and the quadratic statistics such as entropy and correlation of gray level co-occurrence matrix of homogeneous region sample are obtained.The gray mean and mean variance of homogeneous region samples are calculated. The quadratic statistics and mean square deviation of gray level co-occurrence matrix are formed into a feature vector, and the obtained eigenvectors are clustered into hierarchical clusters, and the results of homogeneous region segmentation are obtained.Good segmentation effect has been achieved.The features of the structural region are generally boundary, line target and independent target. Therefore, the watershed method is used to segment the structural region, and the segmentation results of the aggregation region, homogeneous area and structural region are integrated.Finally, the segmentation result of SAR image is obtained.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN957.52

【相似文献】