Parallax Image Stitching Based On Moving Least Squares

Image stitching generally consists of two phases:image alignment and graph-cut and image blending,among which image alignment is a key step and determines the performance of stitching algorithm directly.Therefore,the existing research work mostly focused on the alignment model,and many image deformation methods have been proposed for the image alignment.Image alignment model based on a global transformation is the mainstream of the current stitching method,but most of them have the problem of misalignment of matched feature points,so that they are poor in stitching the large parallax image.The performance of alignment model depends on the correctness of matched feature points and the alignment accuracy of corresponding feature points after image deformation.In order to efficiently remove wrong matches,this paper proposes a new SIFT feature points screening technique based on RANSAC and distance similarity.In the phase of image alignment,we first utilize moving least squares to deform the image so as to ensure the alignment accuracy of every matched feature points,and then present an anti-distortion method based on weighted least squares to adjust transformed image in order to eliminate the local distortion.After the image alignment,we build a network flow model in the overlapped region between transformed image and reference image.Max-flow/Min-cut algorithm is used to look for a globally optimal stitching seam in the network,and finally blends the image pairs along the seam with a weighted image-blending algorithm.Our method is tested on existing and self-established image databases.The most stitching images are realistic,smooth,and the ghost and distortion can be eliminated effectively.The experiment results demonstrate the stitching algorithm is reliable,robust,and stitching performance is better.