A deep learning dense matching method and system with geometric enhancement constraints on relationships

Abstract

This invention provides a deep learning-based dense matching method and system with relational geometry enhancement constraints. The method includes: acquiring a stereo image pair dataset and preprocessing the dataset; constructing an end-to-end convolutional neural network structure RGE-Net; inputting the disparity-corrected stereo image pairs into RGE-Net; the input stereo image pairs are processed in batches through a feature extractor, an adaptive disparity pyramid construction module, a relation-guided geometric weight generation module, a relation-constrained texture feature optimization module, and a relational geometry-guided disparity optimization module; designing a loss function and tuning the parameters of RGE-Net to determine the optimal network model; and feeding the stereo image pairs to be tested into the optimal network model to obtain the disparity prediction results from the network. The RGE-Net network proposed in this invention can achieve accurate disparity estimation for high-resolution stereo image pairs and has certain reference value for dense matching of high-resolution remote sensing images in complex scenes.

Images