A SAR-SAR Template Matching Based on Autoencoders and Multi-Level and Multi-Rotation Feature-Space Correlation

This work presents an unsupervised Deep Learning (DL)-based method for the matching of SAR images in case of high variability of the data. The method is based on an autoencoder trained in an unsupervised manner, with the SAR image in input and output given at the same time. From this architecture, the encoder is detached and used as feature extractor to obtain a set of feature maps at different depths to be used for the comparison. The correlation is applied to these feature maps to get the point of maximum correspondence among the image and the template to be found. The DL model has been trained with a heterogeneous dataset, with images coming from different SAR sources, in order to make it generalised and to make it work in a wide range of scenarios. The method has been tested both with similar and completely different images, showing good performances even in the most critical case.