一种带有注意力机制的大基线场景端到端单应性估计方法  

An end-to-end homography estimation method for large baseline scenes with an attention mechanism

作　　者：林佐江 曹旭 陈玮 白宇 米博 张学伟 方浩[2] LIN Zuojiang;CAO Xu;CHEN Wei;BAI Yu;MI Bo;ZHANG Xuewei;FANG Hao(China Construction First Group Construction&Development Corporation Limited,Beijing 100102,China;School of Automation,Beijing Institute of Technology,Beijing 100081,China;China Construction Municipal Engineering Corporation Limited,Beijing 102627,China)

机构地区：[1]中建一局集团建设发展有限公司,北京100102 [2]北京理工大学自动化学院,北京100081 [3]中建市政工程有限公司,北京102627

出　　处：《河北科技大学学报》2024年第6期669-682,共14页Journal of Hebei University of Science and Technology

基　　金：国家自然科学基金(62133002)。

摘　　要：针对目前现有单应性估计方法存在的精度不高、对大基线场景与运动模糊场景适应性不强的问题,构建了一种带有注意力机制的大基线场景端到端单应性估计方法,采用无监督学习的方式进行单应性估计。首先,引入SE通道注意力模块,构建带有注意力机制的单应性回归网络层,获得网络对于图像各通道间关联性的学习;其次,构建基于掩膜与感知损失度量的二元无监督损失方式,提高网络感知域范围以及网络对于大基线场景的适应性;最后,构建Homo-COCO合成数据集,采用数据增强使得网络模型对于光照变化与运动模糊具有一定的鲁棒性,获得更强的真实场景泛化能力。经过充分的对比及消融实验表明,该方法在精度指标与场景适应性方面优于现有方法,具有良好的准确性与适应性。本方法可以有效估计图像单应性,为图像拼接、图像校正等计算机视觉后续任务提供准确参数估计。Aiming at the problems of low accuracy and limited adaptability to large baseline scenes and motion blur scenarios in current homography estimation methods,an end-to-end homography estimation methodwith attention mechanism for large baseline scenes was constructed,which utilized unsupervised learning for homography estimation.Firstly,by introducing the SE channel attention module,a homography regression network layer with attention mechanism was constructed,enabling the network to learn the inter-channel correlations of images.Secondly,a binary unsupervised loss construction method based on mask and perceptual loss metrics was proposed to enhance the network′s perception range and adaptability to large baseline scenes.Finally,a Homo-COCO synthetic dataset was created,and data augmentation was used to improve the network model′s robustness to changes in lighting and motion blur,resulting in stronger generalization capabilities in real-world scenes.Extensive comparative and ablation experiments demonstrate that this method outperforms existing methods in terms of accuracy and scene adaptability,showing good precision and adaptability.It can effectively estimate image homography and provide accurate parameter estimation for subsequent computer vision tasks such as image stitching and image correction.

关 键 词：计算机图像处理 单应性估计 无监督学习 注意力机制 数据增强 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]