Multi-temporal remote sensing image registration based on multi-layer feature fusion of deep residual network

Abstract

Image registration is a key technology in remote sensing image processing and application. In the registration of multi-temporal remote sensing images, due to differences in imageing conditions, there are two types of typical anomalies in the image change and relative parallax shift between images, which will affect the registration accuracy. Therefore, this paper proposes an algorithm for multi-temporal remote sensing image registration based on depth residual network. In feature extraction stage, multi-scale descriptors are generated from the advanced convolution information of the trained ResNet50 network layer to improve the quantity and quality of feature point extraction. In the registration stage of point set, the difference of feature is calculated by Bhattacharyya distance, and the mismatched point pairs are eliminated by Random Sampling Consistency Algorithms (RANSAC). Finally, the transformation model of the point set is calculated by using the coordinates of the matching point pairs to achieve accurate registration of multitemporal remote sensing images. The experiment uses image data obtained from Google Earth and Lansat 8 satellites and Baidu Map to test the proposed algorithm, and compares it with two feature-based algorithms (PSO-SIFT and CNN). The experimental results show that the proposed algorithm achieves better multi-temporal remote sensing image registration results.