A novel coarse-to-fine method for registration of multispectral images

• We use projective transformation to register multispectral images of large depth.
• Global information is exploited with local descriptors to build keypoint mappings.
• The edge overlap information is utilized as similarity metric.
• The outliers of keypoint mappings are determined by multiple steps.

Due to non-linear intensity changes between multispectral images, the existed descriptors often yield low matching performance. In order to build reliable keypoint mappings on multispectral images, a novel coarse-to-fine method is designed using projective transformation and the information of edge overlap. The method consists of a coarse process and a fine-tuning process. In the coarse process, initial keypoint mappings are built with the descriptors associated with keypoints and the relative distance constraints are employed on them to remove outliers. In the fine-tuning process, the edge overlap information is utilized as similarity metric and an iterative framework is applied to search correct keypoint mappings. The performance of the proposed is investigated with keypoints extracted by speeded-up robust features. The experiment results show that the proposed method can build more reliable keypoint mappings on multispectral images than existed methods.