摘要
针对影像中重复/弱纹理区域的影像高效匹配问题,提出了一种基于深度孪生神经网络的航空立体影像自适应密集匹配方法。以孪生神经网络为基础,构建一种基于图像块逐像素平移匹配思想的立体影像密集匹配深度神经网络模型,对两条分支网络输入的左右影像提取特征向量并进行内积得到匹配代价体积,以计算各像素点的视差值;利用地物的颜色与纹理特征对视差结果进行自适应优化,以对建筑物边缘、遮挡等区域的误匹配情况进行约束,提高了预测结果的可靠程度,有效避免误匹配。使用公开的立体影像数据集对算法进行验证,实验结果表明,在不依赖GPU的情况下,算法产生的视差图更平滑,匹配精度比传统方法提高了30%以上。
Considering the problem of efficient image matching in image regions with repetitive or weak texture,the paper proposes an adaptive dense matching method for aerial stereo images based on deep Siamese network.A deep neural network model for stereo image dense matching is constructed using the concept of pixel-by-pixel translation matching within the framework of a siamese neural network.The feature vectors extracted from the left and right input images by two branch networks are used to obtain a matching cost volume and calculate the parallax value for each pixel.Subsequently,adaptive optimization of the parallax results is performed using color and texture features of ground objects to constrain mismatches in building edges,occlusions,and other areas,thereby enhancing prediction reliability and effectively avoiding mismatching.Finally,the proposed algorithm is validated using publicly available stereo image datasets.Experimental results demonstrate that our algorithm yields superior and smoother parallax maps compared to classical methods with over 30%improvement in matching accuracy without relying on GPU.
作者
赵立科
张帮
张卡
余冰鑫
王玉军
李旋
宿东
张燕平
ZHAO Like;ZHANG Bang;ZHANG Ka;YU Bingxin;WANG Yujun;LI Xuan;SU Dong;ZHANG Yanping(Geological Survey of Jiangsu Province,Nanjing 210018,China;Natural Resources Satellite Application Technology Center of Jiangsu Province,Nanjing 210018,China;Suzhou Fire Rescue Detachment,Suzhou,Jiangsu 215000,China;Key Laboratory of Virtual Geographic Environment(Nanjing Normal University),Ministry of Education,Nanjing 210023,China;School of Geography,Najing Normal University,Nanjing 210023,China;Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application,Nanjing 210023,China;State Key Laboratory Cultivation Base of Geographical Environment Evolution(Jiangsu Province),Nanjing 210023,China;Jiangsu Quality Supervision and Inspection Station for Surveying and Mapping Products,Nanjing 210023,China)
出处
《测绘科学》
CSCD
北大核心
2024年第5期122-132,共11页
Science of Surveying and Mapping
基金
国家自然科学基金项目(42271342,42071301)
江苏省自然资源科技计划项目(2023048)
江苏高校优势学科建设工程资助项目(164320H116)
作者简介
赵立科(1978—),男,高级工程师,硕士,主要研究方向为测绘地理信息方面的工程应用。E-mail:21331090@qq.com;通信作者:张卡,教授,E-mail:zhangka81@126.com
