A superglue image registration algorithm based on adaptive dynamic attention mechanism

To improve image alignment quality, we propose a SuperGlue-based image alignment algorithm, which is an adaptive dynamic graph attention mechanism. The method first uses a deep learning method (SuperPoint) for feature point extraction to get the position, confidence, and encoding vectors of feature points. Next, a multilayer perceptron is used to embed the position, confidence, and encoding vectors together into a high-dimensional vector, and then the information of neighboring points is aggregated using an adaptive attention mechanism to make its node representation vector unique. Meanwhile, to avoid the performance impairment caused by full connectivity, the information range of aggregation is adaptively decayed according to the number of extracted feature points and the number of iterative layers. As the number of layers increases, the information is continuously passed, the graph is continuously updated, and finally the descriptors with unique properties are obtained. Finally, the matching results are obtained by solving the score matrix. The experimental results show that the image alignment quality of our method is improved over previous image alignment algorithms.