Multi-spectral image level-by-level registration algorithm based on A-KAZE feature

Abstract

Aiming at the problem that he gray level of different spectral images varies greatly and the traditional feature extraction algorithm is difficult to maintain the local precision and edge detail of the image, a multi-channel multi-spectral image registration method based on A-KAZE algorithm. In the registration process, the Fast Explicit Diffusion (FED) numerical analysis framework is used to solve the nonlinear diffusion filter equation, and the nonlinear scale space is constructed. The feature points are obtained by calculating the Hessian matrix of each pixel; The invariant image feature vectors are constructed by the Modified-Local Difference Binary (M-LDB) descriptor. Then, the feature vectors are matched by KNN using Hamming distance, and the mismatched points are eliminated by M-estimator Sample Consensus (MSAC). Finally, the transformation matrix is calculated based on projection transformation model. For multi-channel multi-spectral images, the optimal registration route is calculated by level-by-level registration method, and the image registration is realized by registration strategy and transformation matrix. Multispectral phenological observation data were selected to verify the image registration effect of the algorithm, and compared with SIFT, SURF, KAZE algorithm. Experimental results show that this method can achieve sub-pixel registration accuracy on any two images, and has strong robustness and faster speed.