A General Deep Learning Framework Guided by Sparse Matching for Disparity Estimation in High-Resolution Satellite Stereo Imagery

In the field of photogrammetry and remote sensing, the task of satellite stereo image disparity estimation (SSIDE) has long been recognized as both challenging and important. Currently, deep-learning methods are gaining prominence in the SSIDE domain. However, the inconsistency between stereo images and ground truth makes the fine training and accurate inference of SSIDE networks extremely difficult. Furthermore, the existence of textureless and repeated texture areas in satellite images complicates the execution of end-to-end SSIDE networks, especially in areas with variable illumination conditions. In this study, a sparse matching point-guided disparity estimation (SMP-DE) general framework was introduced to address such concerns. SMP-DE employed sparse matching point-guided data evaluation and distillation (SMP-DED) for fault-tolerant training and ensuring unbiased guidance training as well as reliable reasoning. In addition, SMP-DE executed optimization for the disparity estimation network across various feature spaces by integrating sparse matching point-guided feature contrastive registration (SMP-FCR) and matching cost uniqueness constraint (MCUC) modules. Therefore, SMP-DE can mine homogenous features and model low-entropy matching costs in challenging regions. Experimental results demonstrated that SMP-DE has outstanding disparity estimation accuracy and generalization compared with other advanced methods. Furthermore, the proposed SMP-DED exhibited excellent flexibility and generality, since it can be combined with various disparity estimating networks, giving the networks an accuracy boost on a range of datasets. In summary, SMP-DE provides a novel perspective for end-to-end SSIDE research.