Research On Image Registration And Mosaic Based On SURF

Image Registration, as a key technology in image processing field, is the crucial step of various operative technologies such as image mosaic, stereoscopic vision, change detection, image fusion of different sensors, etc. image mosaics, as a important application of image registration can be applied to various fields such as virtual reality, remote sensing image processing, video compression and transmission.SURF is a new feature extraction algorithm which is proposed recently, it approximates or even outperforms previously proposed schemes with respect to repeatability, distinctiveness, and robustness. And it is much faster than the the other similar ones. We use SURF to achieve image registration: firstly, we use SURF to detect and descript the interest points; secondly, we combine the sign of the trace of Hessian matrix and the ratio of the closest neighbor and second closest neighbor to match the interest points, and we eliminate the wrong matches using RANSAC; thirdly, we use DLT to get the least square solution of the homography matrix; lastly, we achieve image registration through interpolation. After image registration, we do some research on image fusion, at last we achieve image mosaics using SURF-based image registration method.My jop in this paper can be summarized into 3 items:(1)We implemented SURF-based image registration method , and did some analysis on its advantage in the stage of the interest points detection and description. We compare the time consuming of the interest points match between SURF-128 and SURF-64. We prove that the sign of the trace of the Hessian matrix can approve the efficiency of interest points matching.We implemented RANSAC constrained by Epipolar Geometry and RANSAC constrained by Homography matrix respectively. And we did some compare between them. We analysed the influence of SURF-64 on the time consuming of eliminating the wrong matches.(2)We made quantitative analysis on the precision of the SURF-based image registration method, and compare it with other image registration method.We made image registration experiments of remote sensing image acquired at different viewpoints, different sensors and different times.(3)We implemented two image fusion methods:gradual change blending and multi-band blending, and did some compare between them.We used SURF-based image registration method to do image mosaics, and implemented the cylindrical panoramic image at last. We proposed a new method based interest points match to sort the the unordered image sequences. In the process of mosaicing frame by frame, we set a sliding window to avoid repeat detection of the reference image which is getting larger and larger.The experiments indicate that on the premise of keep accuracy and applicability, SURF can improve the computation efficiency of image registration obviously. We prove that SURF-based image registration method has much pratical value in remote sensing image processing through successful registration of various remote sensing image. We believe SURF-based image registration method also has much pratical value in image mosaic because of its successful building of cylindrical panoramic image.