Robust Heterogeneous Model Fitting for Multi-source Image Correspondences
Shuyuan Lin,Feiran Huang,Taotao Lai,Jianhuang Lai,Hanzi Wang,Jian Weng
DOI: https://doi.org/10.1007/s11263-024-02023-9
IF: 13.369
2024-02-24
International Journal of Computer Vision
Abstract:Traditional feature detection and description methods, such as scale-invariant feature transform, are susceptible to nonlinear radiation distortions (NRDs) and geometric distortions (GDs), which in turn generate a large number of outliers or incorrect correspondences. To address this issue, this paper proposes a simple yet effective heterogeneous model fitting (MIMF) for multi-source image correspondences. First, a multi-orientation phase consistency model is constructed, which fuses phase consistency, image amplitude and orientation to detect the correct correspondences of feature points. This model effectively reduces the influence of NRDs. Second, sub-region grids and orientation histograms are exploited to construct the log-polar descriptors with variable-size bins, which are robust to GDs. Finally, a heterogeneous model fitting method is proposed, which can effectively estimate the parameters of the transformation model for alleviating the influence of outliers. Experiments are performed on six public datasets and one constructed dataset containing ten types of multi-source images, and the experimental results show that the proposed MIMF method outperforms several state-of-the-art competing methods in terms of matching performance.
computer science, artificial intelligence