Parallax Image Stitching Based On Multi-level Feature Extraction And Matching

The current image registration and stitching algorithms have some common problems,such as,non-globality of its homography,high computational complexity,low matching accuracy,and ghosting and structure deformation in the stitching result.To solve these problems,a novel parallax image stitching algorithm based on the multi-level feature extraction and matching is presented in this paper.Firstly,we propose a novel parallax image stitching algorithm based on feature blocking to improve calculation efficiency.Secondly,a parallax image matching algorithm based on multi-level feature extraction and matching is presented to improve registration precision.Finally,we design a parallax image local optimization model to eliminate parallax ghosting and shape distortion in the matched image.The main contributions of this paper are as follows:1)We propose a parallax image stitching algorithm based on feature blocking.Firstly,Graph Cut algorithm is used to partition the reference image Ii and the objective image I2 into several different blocks with unique features which are numbering respectively.Secondly,the feature matching blocks pairs between the reference image I1 and the destination image I2 are selected using feature blocking algorithm.Finally,the global homography Hi can also obtained by using the feature blocking algorithm.The experimental results show that our approach can guarantee the globality of the homography,reduce the number of iterations,and improve computational efficiency.2)To improve the multilayer matching framework,a parallax image matching algorithm based on multi-level feature extraction and matching is constructed in this paper to match the parallax images.At first,multi-level feature extraction and matching framework is constructed;Secondly,multi-level features are extracted;At last,the multi-level features are matched.During the multi-level feature extracting,layer-by-layer characteristics are converged from high resolution layer to structure layer to guarantee the accuracy of feature extracting.In the multi-level feature matching,matching results in structural layer are used to guide the matching in high resolution layer under the coarse-to-fine strategy in order to accelerate registration.The comparison experiments show that our approach can improve the registration accuracy.3)We design a parallax image local optimization model to eliminate ghosting and shape distortion in the overlapping region caused by parallax image stitching.At the beginning,the proposed "structure + texture" method is adopted to search the optimal global homography H,which is used to transform objective I2 to get pre-aligned objective image I2.After that,the overlapping regions in the pre-registration image I2 are optimized to get the refined image I2 by combining block chaining constraints with position transform constraints and shape distortion constraints.Last,the bilinear interpolation method is used to get the final panoramic stitching image.The experiments demonstrated that our approach can eliminate ghosting and distortion very well and improve the matching quality very high.