Application of improved A-KAZE algorithm in image registration

Aiming at the problem that local precision and edge details are difficult to preserve in the existing process of image registration, an improved image feature extraction algorithm AKAZE-ILDB (accelerated KAZE-improved local difference binary) is proposed based on the A-KAZE algorithm. First, this algorithm uses nonlinear diffusion filtering equation to construct the image pyramid. The numerical solution is obtained by the fast explicit diffusion (FED) method. The coordinates of the image feature points with subpixel precision are obtained. Then, the invariant image feature vectors are constructed by the improved LDB descriptor. The eigenvectors are matched by KNN (K-nearest neighbor) with Hamming distance. Finally, the spatial mapping parameter matrix is computed based on the affine transformation model to realize image registration. The experimental results show that in terms of registration efficiency, the AKAZE-ILDB algorithm reduces average registration time by 300 ms compared with the original A-KAZE algorithm in the condition of maintaining the same matching accuracy. Meanwhile, the matching accuracy of the same image feature is also improved by 3.7% higher than the A-KAZE algorithm and 29% higher than the traditional feature extraction algorithm SURF (speed up robust feature).