多源遥感影像深度特征融合匹配算法

针对多源遥感影像之间成像机理不同、非线性光谱辐射畸变大以及灰度梯度差异明显等所导致的匹配困难问题,提出深度特征融合匹配算法(Feature Fusion Matching Algorithm, FFM)。(1)通过构建特征图金字塔网络提取影像深度特征,使用特征连接结构将语义丰富的高层特征与定位精确的低层特征互补融合,解决多源影像同名特征难以表征的问题并提高特征向量的定位精度;(2)对原始维度1/8的特征图进行交叉变换来融合自身邻域信息与待匹配影像特征信息,通过计算特征向量间的相似性得分得到初次匹配结果,针对特征稀疏区域,提出滑动窗口自适应得分阈值检测算法来提升匹配效果;(3)将匹配结果映射至亚像素级特征图,在小窗口内计算像素间的匹配概率分布期望值来检校优化匹配结果,提高匹配点对的准确性;(4)使用PROSAC算法对匹配结果进行提纯,有效剔除误匹配的同时最大限度保留正确匹配点。试验选取6对多源遥感影像,将FFM同SuperPoint、SIFT、ContextDesc以及LoFTR算法进行对比,结果表明FFM算法在匹配点正确率、匹配点均方根误差以及分布均匀度等方面远优于其他算法。将FFM匹...

Multi-source Remote Sensing Image Deep Feature Fusion Matching Algorithm

Focusing on the difficulty in image matching caused by different imaging mechanisms and large nonlinear spectral radiation distortion between multi-source remote sensing images, a deep Feature Fusion Matching (FFM) algorithm is proposed in this study. Firstly, the feature pyramid network is constructed to extract image deep features, and the feature connection structure is used to complementarily fuse high-level features with rich semantics and low-level features with accurate positioning, so as to solve the problem of difficult representation of homonymous features in multi-source remote sensing images and improve the positioning accuracy of feature vectors. Secondly, the feature map of the original dimension 1/8 is cross transformed to fuse its own neighborhood information and the feature information of the image to be matched. The first matching result is obtained by calculating the similarity score between the feature vectors. For the sparse feature area, an adaptive score threshold detection algorithm using sliding window is proposed to improve the matching effectiveness for sparse feature regions. Then the matching results are mapped to the sub-pixel feature graph, and the expected value of the matching probability distribution between pixels is calculated in a small window to check and optimize the matching results and improve the accuracy of matching point pairs. Finally, the PROSAC algorithm is used to purify the precise matching results, which can effectively eliminate the false matching and keep the correct matching points to the maximum extent. The experiment selects six pairs of multi-source remote sensing images, and compares FFM with SuperPoint, SIFT, ContextDesc, and LoFTR algorithms. The results show that the FFM algorithm is superior to other algorithms in terms of number of correct matching point pairs, matching point accuracy, matching point root mean square error, and matching point distribution uniformity. The FFM matching results are used for multi-source remote sensing images registration, and the registration efficiency is also greatly improved.