Multimodal Remote Sensing Image Match Based on Structure Tensor Feature Transform

Abstract

Traditional feature matching methods like scale-invariant feature transform (SIFT) tend to work only on single modality images, but become poor on multimodal images due to nonlinear mapping between images. To cope with this problem, this letter proposes to use the invariance of structural information in images to detect and describe feature points, which is called structure tensor feature transform (STFT). First, STFT proposes anisotropic Gaussian directional derivative (AGDD) instead of first-order derivative for edge detection, which improves the ac-curacy and noise reduction ability of edges in multimodal images. Then, structural tensor feature decomposition uses each two edge images to decompose into a linearity map and an orientation map. Finally, sequence linearity maps and orientation maps are fused into a intensity map with rich structural information to obtain feature points and a feature description map. Experiments carried out on six types of multimodal images point out that the STFT can effectively overcome the difference in multimodal images compared to the state-of-the-art methods.