Feature Matching and Position Matching Between Optical and SAR With Local Deep Feature Descriptor

Synthetic aperture radar and optical image registration using local and global feature learning by modality-shared attention network

“…To validate the performance of the proposed MSA-Net, the experiments are performed on a public SAR and optical image dataset SEN1-2 36 and five pairs of SAR and optical images. We compare our method with three handcrafted and three deep learning methods, including OS-SIFT, 16 PCSD, 18 RIFT, 19 HardNet, 21 MatchosNet, 26 and CNet 27 . Section 4.1 describes a data description.…”

“…Hughes et al 23 , 24 explained that pseudo-Siamese networks are better suited for multimodal image matching. MatchosNet 26 and CNet, 27 based on pseudo-Siamese network structure, have achieved average results in SAR and optical image registration. MatchosNet is template matching, which can obtain better results only when there is no large translation or rotation transformation between the two images to be registered.…”

“…Both traditional handcrafted methods [13][14][15][16][17][18][19] and deep learning-based ones have been used to address these issues. [20][21][22][23][24][25][26][27][28] The handcrafted image registration approaches mainly include intensity-based and featurebased methods. Due to the poor similarity of intensity-based descriptors, the former is less effective in multimodal image registration.…”

“…Both traditional handcrafted methods 13 – 19 and deep learning-based ones have been used to address these issues 20 – 28 …”

“…For example, L2-Net, 20 HardNet, 21 and Siamese fully convolutional network (SFcNet) 24 based on Siamese convolutional network have been proposed for single- or multimodal image matching or registration. MatchosNet 26 and CNet 27 based on pseudo-Siamese convolutional networks are proposed for SAR and optical image matching and registration.…”