CIFT: Multi-Modal Image Matching Based on Construction-Invariant Feature Transform

With the continuous development of remote sensing technology, the variety of new remote sensing sensors is increasing day by day. At the same time, the single-mode remote sensing image is increasingly difficult to meet the requirements of geological analysis due to its single imaging band. Therefore, the comprehensive application of multi-source remote sensing image has become a research hotspot in the field of remote sensing. Image registration is the prerequisite for the comprehensive application of multi-mode remote sensing image, and the accuracy of image registration directly affects the accuracy of subsequent data analysis results. However, due to the different imaging bands and imaging mechanisms of different sensors, there are large radiation differences between the registration image pairs, and they are affected by noise. To solve this problem, this paper proposes a multimodal image registration method based on the Construction-invariant Feature Transform (CIFT). To reduce the effect of noise on the phase consistency model, anisotropic filtering is applied to the registration image pairs, which can smooth the noise while preserving the structural information of the image. Then, the structural information of the image is extracted using phase consistency, and the high repetition rate feature points are extracted from the phase consistency map. Based on the phase consistency map, the maximum gradient index map (MGIM) is constructed for feature description, which has good robustness against radiation distortion. In the experimental stage, six different remote sensing images including optical, SAR, depth, lidar, infrared and day and night were used to test CIFT and three comparison algorithms (OS-SIFT, SAR-SIFT and PSO-SIFT). Experimental results show that the CIFT method is the only method that can successfully register 6 types of multimodal images, and the registration accuracy is within 3 pixels, and the registration effect is good.