An improved registration and mosaicking method for remote sensing images under geometric and photometric constraints

This paper presents a new automatic and fast satellite image registration approach. In this study, we exploit the invariant associations between different areas of the referenced and sensed images. Mostly existing registration methods are only interested in the extraction of points of interest without taking into account their quality and matching. Our method responds to this problematic because it involves firstly to find a class of separate and independent keypoints which are unchanging to different geometric transformations such as translation, rotation and scaling in addition to photometric constraints and viewpoint adjustments by using the SIFT (Scale Invariant Features Transform) Detector. Then, the matching strategy of SIFT is replaced by a fast k-nearest neighbour algorithm based on the Haar wavelet transform to reduce the space of interest points pair. Finally, to solve the problem of geometric constraints and to estimate the transformation matrix between stereo image pairs, features pairing are filtered using the RANSAC (RANdom Sample Consensus) algorithm. The final warping of the images conforming to the chosen key points is completed by applying a projective transformation and used in a multi-band blending to create a perfect panorama. Experiments on various remote sensing images show the effectiveness and the robustness of our method.