摘要
针对光学-合成孔径雷达(Synthetic Aperture Radar,SAR)异源遥感图像的匹配问题,提出了一种基于图像块描述子的深度学习算法。该方法首先使用图像直方图均衡化算法提升图像质量,在此基础上使用Harris与SIFT算法提取待匹配图像块,再使用提出的深度学习方法计算图像块描述特征向量,最终使用改进的RANSAC算法求解光学-SAR图像变换矩阵。通过80组真实的光学-SAR异源遥感图像进行实验,结果表明该方法相较于无预处理与L2Net算法可以获得更高的匹配数量,并随机选取SEN-12数据集中的四幅图像直观展示了匹配结果。
Aiming at the matching problem of optical-SAR heterogeneous remote sensing images,this paper proposes a deep learning algorithm based on image block descriptors.This method first adopts the image histogram equalization algorithm to improve the image quality,then uses Harris and SIFT algorithms to extract the image block to be matched,and then uses the deep learning method proposed in this paper to calculate the image block description feature vector,and finally uses the improved RANSAC algorithm to solve the optical-SAR image transformation matrix.Experiments with 80 sets of real optical-SAR heterogeneous remote sensing images show that this method can obtain a higher number of matches compared with no preprocessing and L2Net algorithm.And by randomly selecting four images in the SEN-12 data set,the matching result is visually displayed.
作者
阴俊恺
张正强
吴震
曾兵
李斌
YIN Junkai;ZHANG Zhengqiang;WU Zhen;ZENG Bing;LI Bin(Chengdu 30kaitian Communication Industry Co.,Ltd.,Chengdu Sichuan 610041,China)
出处
《通信技术》
2021年第2期344-351,共8页
Communications Technology
关键词
异源遥感匹配
图像块描述子
深度学习
卷积神经网络
heterogeneous remote sensing matching
image block descriptor
deep learning
onvolutional neural network
