基于多尺度特征融合与重要性排序损失的遥感图像匹配网络  

Remote Sensing Image-Matching Network Based on Multiscale Feature Fusion and Importance Ranking Loss

作　　者：陈鹏 包倍源 陈旭 Chen Peng;Bao Beiyuan;Chen Xu(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300401

出　　处：《激光与光电子学进展》2024年第18期266-274,共9页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(U20A20201)。

摘　　要：遥感图像匹配是对地观测的基本问题之一,由于遥感图像中地表信息复杂、尺度多样,往往会对遥感图像匹配造成困难。为此,提出基于多尺度特征融合与重要性排序损失的遥感图像匹配网络。该网络由关键点检测网络和特征描述子提取网络两部分构成。在关键点检测网络中,设计了基于特征金字塔的多层卷积结构,使多尺度特征融合在不同网络层级上实现,并在同一层级中利用多个卷积核逐渐扩大感受野,从而更充分地捕获遥感图像中的多尺度信息。同时,利用CBAM对关键点检测网络的响应图进行聚合,以检测出具有显著得分的关键点。使用分数损失和图像块损失对关键点检测网络进行优化,使用描述子损失对特征描述子提取网络进行优化,并专门设计了分数重要性排序损失函数、描述子重要性排序损失函数,以及基于邻居掩码的描述子损失函数,以保证用于遥感图像匹配的关键点、描述子、图像块具有较高的可重复性、可区分性,从而提高遥感图像匹配的准确性。收集大量遥感图像,通过单应性变换构建遥感图像匹配数据集,并利用该数据集对所提网络模型的性能进行实验验证,发现相比传统图像匹配方法或是其他端到端的深度学习图像匹配方法,所提网络模型在遥感图像匹配中均具有明显优势。Remote sensing image matching is one of the fundamental challenges in earth observation.The complexity and diversity of surface information in remote sensing images often pose difficulties for image matching.To overcome these difficulties,a remote sensing image-matching network based on multiscale feature fusion and importance ranking loss is proposed.This network comprises two parts:a key-point detection network and a feature descriptor extraction network.The key-point detection network has a multilayer convolutional structure based on feature pyramids.This structure is designed to achieve multiscale feature fusion at different network levels.Multiple convolution kernels are used to gradually expand receptive fields at the same level,thereby fully capturing multiscale information in remote sensing images.Furthermore,CBAM is used to aggregate the response graph of the key-point detection network to detect key points with significant scores.The key-point detection network is optimized using the score loss and image block loss,and the feature descriptor sub-extraction network is optimized using the descriptor subloss.The score-importance sorting loss function,descriptor sub-importance sorting loss function,and neighbor mask-based descriptor subloss function are specially designed to ensure that the key points,descriptors,and image blocks used for remote sensing image matching have high repeatability and distinguishability,which improves the accuracy of remote sensing image matching.In this study,many remote sensing images were collected and a remote sensing image-matching dataset was constructed via homography transformation.This dataset was used to experimentally verify the performance of the proposed network model.Compared with traditional image-matching methods or other end-to-end deeplearning image-matching methods,the proposed network model has considerable advantages in remote sensing image matching.

关 键 词：遥感图像 图像匹配 端到端网络 多尺度特征融合 重要性排序损失 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]