残差块改进暹罗网络的遥感图像配准  

Residual Block Improves Remote Sensing Image Registration of SiameseNetwork

作　　者：陈磊 陈颖 李文举 石艳娇 CHEN Lei;CHEN Ying;LI Wen-ju;SHI Yan-jiao(Shanghai Institute of Technology Shanghai 201418,China)

机构地区：[1]上海应用技术大学,上海201418

出　　处：《计算机仿真》2022年第3期224-229,共6页Computer Simulation

基　　金：国家自然科学基金资助项目(61976140);国家自然基金青年基金项目(61806126);上海应用技术大学协同创新基金资助项目(XTCX2018-17)。

摘　　要：针对传统遥感图像配准方法在近红外(NearInfrared,NIR)和可见光(RGB)图像配准上存在识别力不足,准确度低的问题,提出新算法,利用深度神经网络,学习用于NIR和RGB图像块匹配的描述符,并在此基础上构建出一种NIR和RGB遥感图像配准策略。为了更好地学习NIR和RGB图像特征,使用残差块改进密集连接暹罗网络,并增加新颖的交叉熵损失函数减少过拟合,然后通过模版搜索策略生成图像块之间的相似性得分。相似度得分值不仅用于确定对应点位置,还用于配准变换关系的计算,得到配准结果。上述算法在Landsat8中的五个数据集上进行了测试,仿真结果表明:上述算法有效地提高了网络对NIR和RGB遥感图像的识别力和配准的准确度。Traditional image registration methods have insufficient recognition power and low accuracy in near-infrared(NIR)and visible light(RGB)image registration.And so on,the algorithm uses deep neural networks,learns for NIR and better learns NIR and RGB image features,uses residual blocks to improve densely connected spinner networks,and adds a novel cross-entropy loss function to reduce overfitting,and then The similarity between image blocks generated by the template search strategy is outstanding.Registration result.The algorithm was tested on five data sets in Landsat8,and the simulation results show that the algorithm effectively improves the network’s recognition of NIR and RGB images and the accuracy of registration.

关 键 词：图像配准 暹罗网络 模版匹配 遥感图像 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]