基于多路分层稀疏编码的遥感图像场景分类

发布时间：2018-06-16 17:11

本文选题：稀疏编码 + 分层学习　；参考：《西安电子科技大学》2014年硕士论文

【摘要】：在如今多媒体信息技术迅速发展的时代,数字图像越来越多,如何在海量的图像中迅速查找到用户感兴趣的图像或者迅速将图像分门别类便于后续的处理是一个很紧迫的任务。图像场景分类是根据图像内容自动获取图像所属类别的一种技术,已经在模式识别、计算机视觉等领域中得到广泛应用。遥感图像的场景分类作为图像场景分类的一个重要分支,近年来已经对遥感图像的目标检测、图像检索、图像增强等实际问题的研究做出了很大贡献。遥感图像场景分类首先需要从图像中提取特征,然后选择合适的分类器进行分类。所以图像的特征提取很关键,对于分辨度很低的特征,再好的分类器也会失效。遥感场景分类的方法主要分为基于底层特征的方法和基于中层特征的方法两类。基于底层特征的方法不需要识别图像场景中具体的景物,所以相对而言计算复杂度比较低,但是对于图像中比较复杂的场景,底层特征的分类效果很差。这就是底层特征和高层语义之前存在的鸿沟。为解决这种鸿沟,提出了基于中层特征的方法,在底层特征和高层特征之间搭建了桥梁。本文针对图像场景分类主要有以下几个工作:1.介绍了基于多路分层正交匹配追踪(Orthogonal Matching Pursuit,OMP)的半监督遥感图像场景分类方法。不同于传统的基于局部特征描述子的方法,本方法是直接从原始图像块出发学习字典,运用正交匹配追踪的稀疏编码方法、金字塔模型(Spatial Pyramid Matching,SPM)得到整幅图像的特征表示,并结合多路分层学习思想和最大池化方法构建了基于不同大小的图像块的无监督特征学习框架,最后采用半监督的支持矢量机(Semi-Supervised Support Vector Machine,S3VM)进行分类。并将此分类方法扩展到遥感图像的场景检测当中。实验结果表明此算法在遥感图像的场景分类和检测上都能够取得不错的效果。2.提出了基于局部特征描述子和分层稀疏编码的遥感图像场景分类方法。该方法改变传统特征包模型单尺度单层的学习模式,在多个尺度的局部图像块上提取SIFT(Scale invariant feature transform)和LBP(Local Binary Patterns)局部特征描述子,并根据不同尺度的局部特征进行分层稀疏编码,最后将不同尺度下学习的图像特征联合,再用SVM(Support Vector Machine)进行分类。该算法相对传统的基于SIFT和LBP特征的场景分类方法,在遥感图像场景分类中正确率提高了很多。3.提出了基于多路分层正交匹配追踪的半监督图像场景分类方法的MATLAB(MATrix LAboratory)多核并行加速算法。原算法中对图像密集采样、编码、池化等操作均是相同算法对不同数据的独立处理过程,实验的大数据量给参数优化造成很大的困难。本文运用MATLAB多核并行平台将这些计算过程相互独立的算法设计为并行结构。对比实验证明该并行算法大大降低了时间复杂度,解决了优化参数中的难题。
[Abstract]:Nowadays, with the rapid development of multimedia information technology, more and more digital images are available. It is an urgent task how to quickly find the images of interest to the users in the massive images or to classify the images quickly to facilitate the subsequent processing. Image scene classification is a kind of technology which can automatically obtain the category of image according to the content of image. It has been widely used in the fields of pattern recognition computer vision and so on. As an important branch of image scene classification, scene classification of remote sensing images has made great contributions to the research of target detection, image retrieval and image enhancement in recent years. The scene classification of remote sensing images needs to extract features from the image and then select the appropriate classifier for classification. So the feature extraction of image is very important, and for the feature with low resolution, the better classifier will fail. The methods of remote sensing scene classification are divided into two categories: the method based on the bottom feature and the method based on the middle feature. The method based on the underlying feature does not need to recognize the specific scene in the image scene, so the computational complexity is relatively low, but for the more complex scene in the image, the classification effect of the underlying feature is very poor. This is the gap between underlying features and high-level semantics. In order to solve this gap, a method based on middle level feature is proposed, which builds a bridge between the bottom feature and the high level feature. In this paper, the image scene classification has the following work: 1. A semi-supervised remote sensing image scene classification method based on multi-channel hierarchical orthogonal matching tracking orthogonal matching pursuit (OMP) is introduced. Different from the traditional method based on local feature descriptors, this method is to learn the dictionary directly from the original image block, using the sparse coding method of orthogonal matching tracing, and the pyramid model is used to obtain the feature representation of the whole image. The unsupervised feature learning framework based on different size image blocks is constructed based on the idea of multi-path hierarchical learning and the maximum pool method. Finally, semi-supervised support vector machine Semi-Supervised support Vector Machine (S3VM) is used for classification. The classification method is extended to the scene detection of remote sensing images. Experimental results show that the algorithm can achieve good results in scene classification and detection of remote sensing images. A method of remote sensing image scene classification based on local feature descriptor and hierarchical sparse coding is proposed. This method changes the learning mode of single scale and single layer of traditional feature packet model, extracts local feature descriptors of sift scale invariant feature transform) and LBP local binary patterns on local image blocks of multiple scales, and performs layered sparse coding according to local features of different scales. Finally, the image features of different scales are combined and classified by SVM support Vector Machine. Compared with the traditional scene classification method based on sift and LBP, the algorithm improves the accuracy of scene classification in remote sensing images by a lot of .3. A multi-core parallel acceleration algorithm for semi-supervised image scene classification based on multi-channel hierarchical orthogonal matching tracking is proposed in this paper. In the original algorithm, the operations such as dense image sampling, coding and pool processing are all independent processing processes of different data by the same algorithm, and the large amount of experimental data makes parameter optimization very difficult. In this paper, MATLAB multi-core parallel platform is used to design these algorithms which are independent of each other as parallel structure. The comparison experiments show that the parallel algorithm greatly reduces the time complexity and solves the problem of optimizing parameters.
【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】：2014
【分类号】：TP751

【参考文献】