QQ客服 官方微博 反馈留言

题名	基于深层神经网络的遥感图像目标检测
作者	陈雪云
学位类别	工学博士
答辩日期	2014-05-28
授予单位	中国科学院大学
授予地点	中国科学院自动化研究所
导师	刘成林
关键词	遥感图像 目标检测 曲率方向直方图 深层卷积神经网络 Remote Sensing Images Object Detection Histogram of Oriented Curvature Deep Convolutional Neural Networks
其他题名	Target Detection from Remote-Sensing Images Using Deep Neural Networks
学位专业	模式识别与智能系统
中文摘要	目前最先进的卫星遥感成像已经接近0.1米的分辨率，然而卫星快速地扫描陆地和海洋时，会产生海量的图像数据，依靠人眼 进行判读很难做到实时和无遗漏。因此，遥感图像目标自动检测方法具有重要的意义。针对已有遥感图像目标检测方法在搜索定位、特征提取和分类方面的不足，本文提出了几种有效的基于深层神经网络的目标检测方法，并建立了两个大型的分别用于飞机和车辆检测的数据库，在实验中验证了提出方法的优越性。本文的主要创新点如下： 1、为了提高基于滑动窗的目标检测的速度，提出了基于边界光滑性的阈值计算技术、基于积分投影的目标定向技术和一种基于多阈值分割图像的多尺度的滑动窗技术，在遥感图像目标检测实验中，搜索效率比定步长滑动窗方法提高12-20倍。 2、提出了曲率方向直方图（HOC）特征用于遥感图像目标检测。和主流的方向梯度直方图（HOG）和局部二值化模式（LBP）相比，HOC具有更强的稳定性和识别能力。在LFW人脸识别，MNIST手写数字识别，车辆检测三个数据库的实验中，HOC+LBP的性能显著地优于 HOG+LBP，HOG，LBP等其他特征。为了减少非线性SVM的训练时间，提出了一种基于最优间隔(margin)的核参数估计算法。理论分析和实验证明该算法能够显著地缩小最优核参数的范围。 3. 提出了一种混合深层卷积神经网络方法（HDNN），和深层卷积神经网络（DNN）相比，HDNN能够提取多尺度特征，对目标的尺度变化具有更强的适应能力， 能够提取更细微的特征。车辆检测实验表明，HDNN的虚警率比DNN要降低40\%。在MNIST数据库上，HDNN打破了DNN的错误率记录。 4.提出了一种平行深层卷积神经网络方法（PDNN），把多个DNN进行并行融合。和DNN相比，PDNN能够融合不同类的图像数据，避免不同类图像数据在特征提取过程中的互相干扰，从而提高了目标检测能力。车辆检测实验证明，同等规模下，两个分支的PDNN的性能高于两个独立DNN的叠加。
英文摘要	Now the most advanced satellite has achieved an accuracy of near 0.1-meter per pixel. Therefore, when the satellite scan over the lands and oceans in a swift speed, a huge amount of image data will be produced, which can not be read and interpreted by human in a short time, and developing an automatical and efficient target detection method becomes an imperative and essential task. Aimed at the deficiencies of the existing methods in object search and location, feature extraction and classification in remote sensing images, we proposed several efficient methods base on deep neural networks, and build two large databases for detection of aircraft and vehicle respectively. The Experiments verify the significance of our new methods. The main contributes of this paper are: 1. For improving the searching efficiency, we proposed a new thresholding method based on edge smooth property, a orientation computing method based on projection curve and a new sliding window method based on multi-thresholding images of gray scale or gradient. The new method has a 12-20 higher efficiency than the traditional sliding window method. 2. A new histogram of oriented curvatures (HOC) feature is proposed for target detection. Compared with histogram of oriented gradient(HOG) and local binary pattern(LBP), HOC showed more robustness and discriminability. In experiments on LFW，MNIST and vehicle detection database, HOC+LBP outperforms HOG+LBP, HOG, LBP with significant margins. In order to reduce the training time of Nonlinear SVM， we presented an algorithm to estimate the optimal kernel parameter based on the optimal margin. The theoretical analysis and experiments show the validness of the algorithm. 3. A hybrid deep convolutional neural network (HDNN) is proposed. Compared with deep convolutional neural network (DNN), HDNN is capable of extracting multi-scale features, more powerful ability to adapt to scale varieties of object. HDNN can extract more subtle features than DNN. The experiments on vehicle database showed that HDNN deduced the false alarm rate of DNN by 40\%. In the database of MNIST, HDNN gets a new record. 4. A parallel deep convolutional neural network (PDNN) is proposed, which was capable of fusing different types of features by different branches. PDNN can avoid the interfere of different types of images in the feature extraction process. This made PDNN improve the performance of the object detector. The vehicle detection experiment showed that PDNN...
语种	中文
其他标识符	201018014628031
内容类型	学位论文
源URL	[http://ir.ia.ac.cn/handle/173211/6633]
专题	毕业生_博士学位论文
推荐引用方式 GB/T 7714	陈雪云. 基于深层神经网络的遥感图像目标检测[D]. 中国科学院自动化研究所. 中国科学院大学. 2014.

个性服务

查看访问统计

相关权益政策

暂无数据

收藏/分享

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。