基于双重注意力机制的CycleGAN海岸线自动提取方法  

CycleGAN Coastline Automatic Extraction Method Based on Dual Attention Mechanism

作　　者：卢鹏[1] 张娜 邹国良[1] 王振华[1] 郑宗生[1] Lu Peng;Zhang Na;Zou Guoliang;Wang Zhenhua;Zheng Zongsheng(College of Information,Shanghai Ocean University,Shanghai 201306,China)

机构地区：[1]上海海洋大学信息学院,上海201306

出　　处：《激光与光电子学进展》2022年第12期82-92,共11页Laser & Optoelectronics Progress

基　　金：上海市地方能力建设项目(19050502100);上海海洋大学科技发展专项(A2-2006-20-200211)。

摘　　要：将遥感图像进行像素级海陆分割是海岸线提取的一项基础性工作。由于海岸线的动态变化,获取精准的海岸线标记数据集比较困难,为此采用Google Aerial photo-Maps配对样本,在对Google Maps进行海陆二值化处理后构建了新的配对数据集。针对新数据集样本较少问题,在循环生成对抗网络(CycleGAN)模型的基础上,提出了基于双重注意力机制的DAM-CycleGAN。新模型全面考虑遥感图像和海陆二值化图像之间的结构相似性,改进了循环一致性损失,并设计通道注意力模块和空间注意力模块来凸显显著性特征和区域,以增强模型在小样本训练下的特征学习能力。在均方误差、平均像素精度和平均交并比(MIoU)三个评价指标上,与全卷积神经网络模型、DeepLab模型在多个规模数据集训练下的实验结果对比,改进模型转换的海陆二值化图像与真值图像更加吻合,MIoU值分别至少提高7%、6%以上,验证了所提方法的有效性和可行性。The pixel-level sea-land segmentation of remote sensing images is a basic work for coastline extraction. Owing to the dynamic changes in the coastline, obtaining accurate coastline marker datasets is difficult. In this study, Google Aerial Photo-Maps-paired samples were used to construct a paired dataset after the sea-land binarization processing of Google Maps. Thus, we proposed the dual attention mechanism-cycle generative adversarial network(CycleGAN) based on the CycleGAN model to solve the problem of fewer samples in the new dataset. The new model fully considers the structural similarity between remote sensing images and sealand binarized images, improves cycle consistency loss, and designs both channel and spatial attention modules to highlight salient features and regions to enhance the model’s performance in small feature learning ability under sample training. Furthermore, we applied three evaluation indicators, i. e., mean square error, mean pixel accuracy, and mean intersection over union(MIoU), and compared our experimental results to those of the full convolutional neural network and DeepLab models under multiple-scale dataset training. Results show that the improved model conversion of the sea-land binarized images is more consistent with the true value images and the MIoU values are increased by at least 7% and 6%, respectively, verifying the effectiveness and feasibility of the proposed method.

关 键 词：图像处理 遥感 循环生成对抗网络 注意力机制 循环一致性损失 小样本 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]