摘要
单应性矩阵在图像拼接、3D重建以及同步定位与建图(Simultaneous Localization and Mapping,SLAM)等方面发挥着重要作用。传统的单应性矩阵估计方法依赖于特征点的提取,基于深度学习的单应性矩阵估计模型可以有效解决该问题。但是,当图像不是平面或者仅仅因为旋转、平移放缩而不同时,仅仅依靠一个全局单应性矩阵进行图像拼接会出现重影情况。针对这种情况,笔者提出基于深度学习的多映射无监督单应性矩阵估计模型,通过引入计算机图形学中的网格概念,针对每个网格估计出一个与之对应的局部单应性矩阵,从而适应图像中的不同深度平面。
Homography matrix plays an important role in image mosaic, 3D reconstruction and slam. The traditional homography matrix estimation method depends on the extraction of feature points. The homography matrix estimation model based on deep learning can effectively solve this problem. However, when the image is not planar or is different only because of rotation, translation and zooming, ghosting will occur when only relying on a global homography matrix for image stitching. In view of this situation, the author proposes a multi mapping unsupervised homography matrix estimation model based on deep learning. By introducing the concept of grid in computer graphics, a corresponding local homography matrix is estimated for each grid, so as to adapt to different depth planes in the image.
作者
和青
潘志松
罗建欣
HE Qing;PAN Zhisong;LUO Jianxin(School of Command and Control Engineering,Army Engineering University of PLA,Nanjing Jiangsu 210007,China)
出处
《信息与电脑》
2022年第6期54-57,共4页
Information & Computer
关键词
图像拼接
深度学习
单应性
卷积神经网络
image mosaic
deep learning
homography
convolution neural network
