Deep Dense Projection Estimator for Images With Multi-Geometric Structures.

In this letter, we propose a novel dense projection estimator called deep dense projection (Deep-DP) for 2-D image matching. Deep-DP addresses the challenge of nonsingle projection in ultralow-altitude remote sensing scenes caused by moving viewpoints and parallax of 3-D views. Deep-DP is based on affine transformation and consists of 23 cascaded layers, with specific design ideas to achieve reliable and accurate results. These design ideas include downscaled random sampling, random/semi-random sampling, abnormal matches filtering, and structural parallax suppression. To accommodate the cascade estimators, we have designed a GPU parallel computing framework that ensures high robustness, accuracy, real-time performance, and a low failure rate for Deep-DP. Our experiments demonstrate that Deep-DP outperforms the state-of-the-art single-model robust estimators by having fewer faults and higher fitting power. Additionally, Deep-DP exhibits superior accuracy compared to the state-of-the-art multistructure estimators.