A Novel Self-supervised Affine-invariant Feature Learning Method for Remote Sensing Image Matching

Robust image matching is a prerequisite for many common tasks in the field of remote sensing. Because of the lack of human-labeled feature matching datasets, it is a challenging task to develop techniques for accurate and robust image feature matching in remote sensing. For this problem, this paper proposes a self-supervised affine-invariant feature learning algorithm for remote sensing imagery. The algorithm mainly consists of three steps, firstly the pretrained deep convolution neural network is trained to obtain high-level semantic features from remote sensing imagery pairs, next affine-invariant features are extracted by means of self-attentional graph neural network, which is trained using a self-supervised paradigm. Finally, feature matching is achieved utilizing mutual nearest neighbor criteria. The experimental results demonstrate that the proposed affine-invariant feature learning method is able to achieve satisfied performance and robust to significant geometric distortion of images.