Deep homography convolutional neural network for fast and accurate image registration
Xingzi Liu, Lin Luo, Chaoyong Peng, Jinlong Li · Twelfth International Conference on Information Optics and Photonics · 2021
In this paper, a deep convolution neural network for image registration using homography transformation is proposed to improve the speed and accuracy of image registration. The four-corner homography parameterization is carried out by randomly clipping and perturbing the image block, and then the mapping from one image to another is completed to form the homography image registration dataset. In this network architecture, the homography matrix is obtained by returning the mean square error to the corner variables of the local region. In the preprocessing stage, the image is equalized by the histogram and the feature is magnified. The trained homography matrix is used for the affine transformation of the registered image to verify the effectiveness of the model. We test the dataset of homography image registration and experiment on various noises and various image enhancement effects. We also compare several traditional algorithms. The results prove that the accuracy of the proposed model is state-of-art. The processing speed of a single image is only 0.28 seconds, which has strong noise adaptability and the best performance.