润色论文网提供毕业论文写作、职称论文写作、职称论文发表及期刊论文发表服务!QQ:526140740

电话:18796337551
当前位置:润色论文网系统论文 → 文章正文

基于深度学习的遥感图像特定目标检测算法研究

作者:润色论文网  来源:www.runselw.com  发布时间:2019/10/8 10:29:39  

摘要:近年来,随着遥感成像技术水平的不断提升,遥感图像中的目标检测也受到越来越多研究人员的关注。其中,舰船与飞机作为国家重要的战略资源,在军事以及民用领域有着极大的影响力,因此遥感图像舰船与飞机目标检测成为科研工作者重点研究的对象。但是,由于遥感图像覆盖范围广、背景复杂度高以及成像视角独特使得遥感图像目标检测具有极高的挑战性。近年来,国内外研究人员针对遥感图像舰船与飞机检测问题提出了许多算法,本文在现有算法的基础上,深入研究Faster RCNN目标检测算法的原理,并针对遥感图像舰船与飞机检测任务的难点将算法进一步优化。具体工作内容主要包括:

1. 针对CCCV2017发布的遥感图像数据集背景复杂度高,目标检测易受周围物体干扰问题,本文提出使用深度残差网络提取图像表达能力更强的深度特征。实验结果表明该方法可以有效提高遥感图像目标检测的精度。

2. 本文深入研究FasterRCNN算法区域建议网络生成目标候选区域的原理,并根据遥感图像数据集中目标的特点,对算法中anchor比例进行优化。另外,针对训练数据集中正负样本不平衡问题,本文提出将在线困难样本挖掘机制融入Faster RCNN。实验发现该方法可以进一步提高算法性能。

4. 针对数据集中高分辨率遥感图像训练困难以及在测试阶段小目标难以检测的问题,本文采用滑动窗口的策略对高分辨率图像进行切分,并将切分后的子图输入到Faster RCNN算法进行训练以及检测。实验表明,该方法可以在一定程度上加大训练集的数据量,同时完成高分辨率遥感图像中小目标的检测。

5. 本文将改进后的算法分别与初始的FasterRCNN以及YOLOv2算法进行对比。在CCCV2017遥感图像数据集上评估表明,改进后的算法极大提高了初始的Faster RCNN算法性能,在测试集上召回率达到82.55%,精确率达到73.74%。算法最后对CCCV2017数据集中飞机部分进行单独训练与测试,并将训练后的模型在NWPU VHR-10、NWPU-RESISC45以及UCAS-AOD遥感图像飞机数据集上进行测试,实验结果表明改进后的算法具有良好的鲁棒性和泛化能力。

In recent years, with the improvement ofremote sensing imaging technology, object detection in remote sensing imageshas attracted more and more researchers' attention. Among them, ship andairplane, as very important strategic resources of the country, have greatinfluence in the military and civil fields, so ship and airplane detection in remotesensing images have become the focus of scientific research. However, objectdetection in remote sensing images is extremely challenging due to its widecoverage, high background complexity and unique imaging perspective. In recentyears, researchers at home and abroad have proposed many algorithms for shipand airplane detection in remote sensing images. Based on the existingalgorithms, this paper deeply studied the principle of Faster RCNN objectdetection algorithm, and further optimized the algorithm for the difficultiesin the detection of ship and airplane in remote sensing images. The specificwork mainly includes:

1. The remote sensing image datasetreleased by CCCV2017 has high background complexity and object detection issusceptible to interference from surrounding objects. This paper proposes touse deep residual network to extract depth features with stronger imageexpression ability. Experimental results show that this method can effectivelyimprove the precision of remote sensing image object detection.

2. In this paper, the principle for FasterRCNN algorithm region proposal network to generate object candidate regions isstudied in depth, and the anchor ratio in the algorithm is optimized accordingto the characteristics of objects in remote sensing images. In addition, forthe problem of unbalanced positive and negative samples in the trainingdataset, this paper proposed to integrate the online hard example miningmechanism into the Faster RCNN. Experiments show that this method can furtherimprove the performance of the algorithm.

3. Aiming at the difficulty in traininghigh-resolution remote sensing images in the dataset and the difficulty indetecting small objects in the test stage, this paper adopted the strategy ofsliding window to segment high-resolution images and input the segmentedsubgraphs into Faster RCNN algorithm for training and detection. Experimentsshow that this method can increase the data volume of training set to a certainextent and complete the detection of small objects in high-resolution remotesensing images.

4. In this paper, the improved algorithmwas compared with the initial Faster RCNN and YOLOv2 algorithm respectively.The evaluation on the CCCV2017 dataset shows that the improved algorithm hasgreatly improved the performance of the initial Faster RCNN algorithm, and therecall reached 82.55% and the precision reached 73.74% on the test set.Finally, the algorithm conducts separate training and test on the airplane partof CCCV2017 dataset, and tests the trained model on NWPU VHR-10, NWPU-RESISC45and UCAS-AOD remote sensing image airplane datasets, the experimental results showthat the improved algorithm has good robustness and generalization ability.

关键词:目标检测;FasterRCNN;残差网络;在线困难样本挖掘;滑动窗口

object detection;FasterRCNN;residual network;onlinehard example mining;sliding window

联系方式

客服QQ 526140740
客服热线18796337551
网站地址 www.runselw.com
郑重承诺 原创,包修改,包通过!
润色论文网真诚欢迎新老客户的光临与惠顾!