A rotation matching method for heterogeneous images based on invariant feature translation

Aiming at the problem of heterogenous image matching failure under large angle rotation, a heterogenous image rotation matching method based on invariant feature migration is proposed. First, based on the idea of image style transformation, a GCI-CycleGAN generative adversarial network is constructed to transform satellite images to the infrared images. Then, the HR-LightGlue rotation matching algorithm based on hierarchical reinforcement strategy was designed. SuperPoint is utilized to detect and describe the keypoints before LightGlue is utilized for coarse matching. The images are rotated according to the inter-image transformation relations obtained from coarse matching and exact matching is performed. Finally, the transformation relations obtained from fine matching are mapped to the original image pairs to obtain the final matching results. The invariant feature translation experiments and matching experiments show that the GCI-CycleGAN can effectively improve the quality of invariant feature translation. In addition, when combined with the HR-LightGlue matching method, the matching accuracy is effectively improved.