高分辨率遥感图像深度学习舰船检测技术研究

发布时间：2018-04-22 02:39

  本文选题：高分辨 + 特征提取　； 参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：舰船检测一直是国家海岸带安全的传统任务,我国具有很长的海岸线,通过遥感手段对海岸线进行监控布防,可以大大提高近海地区的防御能力,以及海运的管理、调度能力。目前,随着光学分辨能力的提高,遥感图像的空间分辨能力获得了质的飞升。空间分辨率的提升不仅仅是信息量的提升,同时带来了复杂的处理问题,由于图像细节更加丰富,纹理更加复杂,对图像的分析变得更加困难,通过肉眼获取舰船目标的成本越来越高。而人工设计的特征检测器效果也不尽人意,迫切需要一种合适舰船检测的方法来有效利用大量的数据从而获得更加精确的检测效果。本文从高分辨率遥感图像的多尺度特征出发,研究了深度学习舰船目标检测方法,包括三部分:基于定位框的层次化卷积神经网络方法,结合超像素的超像素卷积神经网络方法,以及进一步的目标分割的方法,具体内容如下:首先,本文针对一般卷积神经网络目标检测方法没有有效利用各层特征以及多层卷积神经网络由于平移不变性带来的定位精度低的缺点。研究了不同层特征对舰船检测的影响,构建了一个定位精度高的单层卷积神经网络,并应用迁移学习的方法构建了一个可以精确区分背景与目标的多层卷积神经网络。在分析了两种网络的优势与劣势的基础上,利用强化学习的思想,结合单层卷积网络与多层卷积网络的优点提出了一种层次化的多尺度深度学习检测方法,获得比标准卷积神经网络更加精确的定位精度以及检测结果。之后,本文针对多尺度滑动窗口计算量大、难以实施的缺点,研究了结合超像素的卷积神经网络目标检测方法,对当前深度学习目标检测主流的对象预提取方法进行分析,研究了通过单一超像素方法获取舰船目标的待选对象的方法,并使用卷积神经网络获得更准确的检测结果。该方法同时具有有监督方法的准确性以及无监督方法的快速性,在不损失图像分辨率的前提下,实现了一种相对精度较高且快速的检测方法。最后,研究了目标分割(像素级检测)方法,采用逐像素取窗的方法获得对图像中每个像素临近像素的准确描述,并区分该像素是否落在目标内,以达到对舰船目标进行分割的目的。并研究了使用超像素中心对区域进行表示的方法,提出了结合超像素表示方法与卷积神经网络方法对舰船目标进行快速的像素级检测的方法,实现了了舰船目标快速的分割提取。
[Abstract]:Ship detection has always been a traditional task of national coastal zone security. Our country has a very long coastline, remote sensing to monitor and control the coastline, can greatly improve the defense capability of offshore areas, as well as the management of maritime transport, scheduling capacity. At present, with the improvement of optical resolution, the spatial resolution of remote sensing images has been improved. The enhancement of spatial resolution is not only the improvement of information content, but also brings about complex processing problems. Because the details of the image are more abundant and the texture is more complex, the analysis of the image becomes more difficult. It is becoming more and more expensive to obtain ship targets with the naked eye. However, the effect of the artificial feature detector is not satisfactory, it is urgent to use a suitable method of ship detection to effectively use a large number of data to obtain a more accurate detection effect. Based on the multi-scale features of high-resolution remote sensing images, this paper studies the method of deep-learning ship target detection, which consists of three parts: hierarchical convolution neural network method based on location frame. The hyperpixel convolution neural network method combined with super-pixel, as well as the method of further target segmentation, are as follows: first of all, In this paper, the general target detection method based on convolutional neural networks does not utilize the features of each layer effectively and the localization accuracy of multi-layer convolutional neural networks is low due to the translation invariance. In this paper, the influence of different layer features on ship detection is studied, a single-layer convolution neural network with high positioning accuracy is constructed, and a multi-layer convolution neural network is constructed by using migration learning method to accurately distinguish the background from the target. Based on the analysis of the advantages and disadvantages of the two networks, a hierarchical multi-scale depth learning detection method is proposed by combining the advantages of single-layer convolution network and multi-layer convolutional network by using the idea of reinforcement learning. Compared with the standard convolution neural network, the location accuracy and detection results are obtained. Then, aiming at the shortcomings of multi-scale sliding window, which is difficult to implement, this paper studies the convolution neural network target detection method combined with super-pixel, and analyzes the main method of object pre-extraction in depth learning target detection. In this paper, a single super-pixel method is studied to obtain the target to be selected from a ship target, and a more accurate detection result is obtained by using convolution neural network. This method has the accuracy of the supervised method and the rapidity of the unsupervised method. Without losing the resolution of the image, a high accuracy and fast detection method is realized. Finally, the method of target segmentation (pixel level detection) is studied. The method of pixel by pixel window is used to obtain the accurate description of each pixel adjacent to the pixel in the image, and to distinguish whether the pixel falls in the target or not. In order to achieve the purpose of ship target segmentation. The method of using the super-pixel center to represent the region is studied, and the method of fast pixel level detection of ship target is proposed by combining the super-pixel representation method and convolution neural network method. The fast segmentation and extraction of ship target is realized.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751;TP18

【参考文献】

中国期刊全文数据库 前4条

1 欧阳颖卉;林，

本文编号：1785264