融合上下文特征与密集网络的光学遥感图像配准

Optical remote sensing image registration with fusion of contextual features and densenet

光学遥感图像的地理位置特征复杂多样,且多尺度特征空间信息丰富,在配准过程中难以充分提取图像特征,配准精度较低。针对上述问题,提出融合上下文特征与密集网络的配准模型,通过把位置信息嵌入注意力中加深对位置信息的关注,并集成多个不同内核的深度可分离卷积整合多个不同感受野来聚合丰富的多阶特征语义信息。首先采用融合后的密集网络对图像进行特征信息提取,接着使用双向皮尔逊相关匹配得到双向匹配关系,并通过回归得到的双向参数加权合成最终参数,最后通过仿射变换完成图像配准。实验结果表明,关键点正确估计的比例指标系数为0.05,0.03和0.01情况下,在Aerial-image数据集中分别高达83.9%,60.3%和15.3%,有效提高了光学遥感图像配准精度。

The geographic location features of optical remote sensing images are complex and diverse,and the spatial information of multi-scale features is rich,which makes it difficult to fully extract the image features in the registration process and the registration accuracy is low.To address the above problems,an registration model that fuses contextual features and densenet is proposed,which deepens the attention to location information by embedding it in attention and integrates several different kernels of depth-separable convolution to integrate several different sensory fields to aggregate the rich multi-scale feature semantic information.Firstly,the fused densenet is used to extract the feature information from the image,then the bidirectional matching relationship is obtained using bidirectional Pearson correlation matching,and the final parameters are synthesized by weighting the bidirectional parameters obtained through regression,and finally the image registration is completed by affine transformation.The experimental results show that the proportional index coefficients of key points correctly estimated at 0.05,0.03 and 0.01 are as high as 83.9%,60.3%and 15.3%in the Aerial-image dataset,respectively,which effectively improves the optical remote sensing image registration accuracy.