张天龙^1,2,3, 过杰^1,2,4
1.中国科学院烟台海岸带研究所中国科学院环境过程与生态修复重点实验室, 山东 烟台 264003;2.中国科学院烟台海岸带研究所 山东省海岸带环境过程重点实验室, 山东 烟台 264003;3.中国科学院大学, 北京 100049;4.中国科学院海洋大科学中心, 山东 青岛 266071

摘要:
SAR（synthetic aperture radar）图像溢油暗斑准确识别对海上溢油应急工作具有重要的意义。为减少SAR图像特征提取、特征选择过程中人为因素对溢油检测精度的影响，本文将Faster R-CNN卷积神经网络模型引入SAR图像溢油检测并进行了改进。针对溢油暗斑形状多样及SAR图像背景复杂的特点，选用结构一致且实用性强的VGG16卷积网络获取图像特征，并使用软化非极大值抑制算法（Soft-NMS）进行优化。同时基于相同的数据集，提取常用的SAR图像几何特征、灰度特征和纹理特征，构建反向传播（backpropagation，BP）人工神经网络溢油检测方法并与Faster R-CNN方法进行对比。实验结果表明，基于改进Faster-RCNN模型的溢油检测方法溢油检测率达到0.78，且溢油检测虚警率低于0.25，相比BP人工神经网络溢油检测方法样本识别率、溢油检测率分别提高了4%和5%，溢油虚警率降低了5%。
关键词:SARFaster R-CNN溢油检测BP神经网络
DOI：10.11759/hykx20200406001
分类号:P76
基金项目:国家重点研发计划项目（2017YFC1405600）；国家自然科学基金（42076197）；国家自然科学基金（41576032）

Oil spill detection method for SAR images based on the improved Faster R-CNN model
ZHANG Tian-long^1,2,3, GUO Jie^1,2,4
1.CAS Key Laboratory of Coastal Environmental Processes and Ecological Remediation, Yantai Institute of Coastal Zone Research, Chinese Academy of Sciences (CAS), Yantai, Shandong 264003, China;2.Shandong Key Laboratory of Coastal Environmental Processes, Yantai Institute of Coastal Zone Research, CAS, Yantai, Shandong 264003, China;3.University of Chinese Academy of Sciences, Beijing 100049, China;4.Center for Ocean Mega-Science, CAS, Qingdao 266071, China
Abstract:
Oil spill emergency work needs to detect oil spills accurately in synthetic aperture radar (SAR) images. To reduce the influence of human factors on oil spill detection accuracy in the SAR image feature extraction and selection processes, the Faster R-CNN model is introduced and improved in this study. Because of the various shapes of oil spills and the complex background, the VGG16 convolutional network with consistent structure and strong practicability is selected to obtain the image features. The Soft-NMS algorithm is used to optimize the Faster R-CNN model. On the basis of the same dataset, the most frequently used geometric, gray, and texture features of SAR images were extracted to build the backpropagation (BP) artificial neural network oil spill detection model, which is compared with the method proposed in this study. The experimental results show that the detection rate of the improved Faster R-CNN model is 0.78, and the false alarm rate is lower than 0.25. Compared with the BP artificial neural network method, the identification and detection rates of the improved Faster R-CNN model are increased by 4% and 5%, respectively, and the oil spill false alarm rate is decreased by 5%.
Key words:SARFaster R-CNNoil spill detectionBP neural network





PDF全文下载地址:

点我下载PDF