CASIA OpenIR  > 毕业生  > 博士学位论文
天基可见光图像舰船目标识别
刘子坤1,2
学位类型工学博士
导师杨一平
2017-05-23
学位授予单位中国科学院大学
学位授予地点北京
关键词遥感可见光图像 舰船数据集 舰船识别 深度学习 卷积神经网络
其他摘要

       舰船识别广泛应用于水上交通安全监管、海洋渔业管理以及军事侦察等方面,是天基遥感应用的研究热点之一。舰船识别的数据源主要包括合成孔径雷达(SAR)图像、可见光图像和红外图像等。传统应用中可见光图像一般作为SAR 图像的重要补充。而随着遥感器技术的发展,天基可见光图像空间分辨率逐步提高,提供了越来越丰富的细节信息,使得识别舰船目标类型或其具体型级逐步具备了可行性,将为决策者提供更加精确、直观的识别信息。因此基于可见光图像的舰船精细识别工作越来越重要。

       现有研究工作主要面对舰船检测任务,极少有工作进行舰船类型精细识别。此外,检测任务中海况较好的海面舰船检测效果较好,但复杂背景下尤其是近岸舰船检测则面临性能瓶颈。已有研究工作的问题在于:基于手工设计的多步骤检测方法,可以从部分复杂背景中抽取出舰船,但非常不鲁棒;基于学习的通用方法,难以应对舰船目标的条状、旋转、聚集的特殊性。因此,在舰船识别领域,迫切需要提出新的研究路线和方法。本文分析了国内外可见光舰船识别相关文献,总结已有研究工作的不足后,提出并建立了第一个(据我们调研所知)标准公开遥感可见光舰船数据集,定义了舰船型号识别任务、提出新的评价标准和基准方法,以两种不同的评价方式对比分析识别结果,并推荐以“旋转矩形框”的方式定位和评价舰船识别任务。此后以此为主题,本文提出了高召回率的基于舰船旋转矩形框空间的舰船预检方法,还设计了基于旋转区域的卷积神经网络以用于舰船分类识别,从而证明了以旋转矩形框为基础的工作路线的潜力以及可行性和有效性。主要研究内容包括:

       1、为了促进舰船识别相关研究,本文提出并建立一个标准的可见光遥感图像舰船目标公开数据集(HRSC2016),提出舰船型号识别任务以及定义新的评价标准和基准方法。数据主要来源于具有重要实验参照价值的谷歌地球(可公开访问)的图像数据。数据集不仅包含矩形框标注方式,还包括能更精准定位舰船目标的旋转矩形框方式,并为每个样本提供三层标签:船、种类、型号。数据集还为有“V”形船头的船标注船头位置信息,为测试集的每一张图像提供分割掩膜。此外,数据集提供用于舰船的标注工具和一些辅助实验工具。本文在数据集上分析了一些经典方法。为了定量评估海陆分离任务,提出了新的评价准则:分割拟合度(SF)。本文还为数据集设计了新的用于舰船识别的基准方法,并采用基于矩形框和旋转矩形框两种方式对比评价分析舰船识别结果。

       2、针对复杂背景下舰船难以提取的问题,本文分析舰船自身特点,设计了一个可快速生成高召回率的一定数量的舰船候选区域的方法。该方法构造了一个近似闭合特性的舰船旋转矩形框空间,利用两级级联的线性模型对空间内的潜在候选区域快速打分,之后利用一个二值线性规划模型选出较小数量的高可能性候选区域。此外,针对潜在的在轨处理应用,还设计了一个性能相近的快速版本。实验证明该方法在为每张图像生成5000个候选框时,取得了97:6%的检测率,超过同类对比方法10个百分点。

       3、针对舰船识别的多层分类问题,本文设计了一个基于旋转区域的卷积神经网络(RR-CNN),可用于旋转目标识别,尤其是舰船识别任务。RR-CNN有三个重要特征,即: 旋转感兴趣区域池化层、稳定的旋转矩形框回归模型和基于多任务的非极大值抑制。RR-CNN具有可精确提取旋转区域特征的优势,同时兼具区域共享卷积计算的特性,因此可实现快速计算、精确定位以及更精准分类。此外,RR-CNN的重要组件还可以与经典框架Faster RCNN, SSD等相结合,将来进一步实现端到端的训练等特性。实验表明在HRSC2016三层舰船识别任务上,我们设计的用于舰船识别的RR-CNN模型与经典的Fast R-CNN相比可提高+18.5到+20.0个mAP百分点,与数据集基准方法相比可提高+4.8到+6.1个mAP百分点,其中在L1级舰船检测任务上取得了75:7% 的平均精度值。

;

      Ship recognition is widely used in vessel traffic services, marine fishery management and military reconnaissance, etc., and is one of the hotspots of spacebased remote sensing applications. The data sources include Synthetic aperture radar (SAR) images, optical images and infrared images. Conventional applications in the optical image is generally used as an important complement to SAR images. But with the long term development of the optical remote sensing technique and the progress of spatial resolution, the optical imagery can provide more details which make the accurate recognition possible. Also the results are more intuitive to a human decision maker. Therefore, it is more and more important to recognize ships based on space-based optical images.

     The existing research works mainly face the ship detection task, few works carry on the fine-grained ship recognition. In addition, existing works achieve good performance in images with calm sea background, but face performance bottlenecks on inshore ship detection task in images with complex backgrounds. The hand-crafted step-by-step methods can detect ships in some images with complex backgrounds, but they are not robust. The methods based on learning are difficult to deal with strip-like rotated assembled objects. Therefore, in the field of ship recognition, there is an urgent need to propose new routes and methods. In this paper, we review the relevant existing works of ship recognition in optical images, build the first (to our best knowledge) public remote sensing dataset for ship recognition and define a ship type recognition task. We also propose new criterions for evaluations and new baselines for ship classification. We compare ship recognition results by two different evaluation ways and recommend the “Rotated Bounding Box” way for ship location and result evaluation. Then, taking this as the principle theme, we propose a new ship proposal generation method with high recall based on ship rotated bounding box space and design a rotated region based convolutional neural network for ship classification, by which we prove the potential, possibility and effectiveness of methods based on rotated bounding boxes. The main contents of this thesis are as follows:

    1. In order to facilitate the research of ship recognition, we present a public high-resolution ship dataset called "HRSC2016" that covers not only boundingbox labeling way, but also rotated bounding box way which provides more accurate location information. We annotate every ship with three-level classes including ship, ship category and ship types. We also provide the ship head position for all the ships with "V" shape heads and the segmentation mask for every image in "Test set". Besides, we volunteer a ship annotation tool and some development tools. Given these rich annotations we perform a detailed analysis about some state-of-the-art methods, introduce a novel metric called the separation fitness (SF) which is used for evaluating the performance of the sea-land segmentation task. We also design two new baselines for ship recognition and analyse the results by two different evaluation ways including bounding box and rotated bounding box ways.

    2. To extract ships from complex backgrounds, we propose a fast method to generate a small number of ship proposals with a high recall. We first construct a nearly closed-from ship rotated bounding box space. Then, by fast scoring for each latent candidate in the space using a two-cascaded linear model followed by binary linear programming, we select a small number of highly potential proposals. Moreover, we also propose a fast version of our method for latent applications running on the satellites. Experiments indicate that our method achieves a detection rate of 97:6% when generating 5000 candidates for each image, which is 10 percentage points higher than that of comparison methods.

    3. For hierarchy classification task, we propose a Rotated Region based Convolutional Neural Network (RR-CNN) for rotated object recognition, especially for ship recognition. RR-CNN has three new important components including a rotated region of interest (RRoI) pooling layer, a rotated bounding box regression model and a multi-task based non-maximal suppression (NMS) between different classes. RR-CNN has the advantages of accurate feature extraction for rotated regions and sharing convolution computation, which are important supports for fast computation, accurate location and more precise classification. In addition, in the future, we can try to combine RR-CNN with the state-of-the-art frameworks (such as Faster R-CNN and SSD, etc.) to get the good features such as end-to-end training, etc. Experimental results confirm that RR-CNN outperforms Fast R-CNN and the baseline on all the three tasks by +18.5 to +20.0 and +4.8 to +6.1 mAP points respectively and achieves 75:7% average precision value on L1 ship detection task.

语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14671
专题毕业生_博士学位论文
作者单位1.Institute of Automation Chinese Academy of Sciences, 95 Zhongguancun East Road, 100190, Beijing, China
2.University of Chinese Academy of Sciences, 80 Zhongguancun East Road, 100190, Beijing, China
推荐引用方式
GB/T 7714
刘子坤. 天基可见光图像舰船目标识别[D]. 北京. 中国科学院大学,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
毕业论文-刘子坤-v6-final-优化(8946KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[刘子坤]的文章
百度学术
百度学术中相似的文章
[刘子坤]的文章
必应学术
必应学术中相似的文章
[刘子坤]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

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。