Research On Seam Guided Image Stitching Methods

The task of image stitching is to stitch multiple pictures with certain overlapping parts into a seamless wide vision panorama.The image stitching technology has been widely used in panoramic photos generation,medical images processing,virtual reality and other fields.Modern image stitching algorithms are usually based on feature matching technology,the transformation matrix of each image is estimated using the matching feature points in the image pair,and then the image registered using that matrix is fused by certain methods to obtain the stitching result.Traditional image stitching usually requires that the input image has little or even no parallax,image stitching in scenes with large parallax is a challenging problem in the field of image stitching.Several studies in recent years have shown that seam-guided method is effective for handling parallex.A novel seam-guided image stitching algorithm based on feature layer separation is proposed,feature matching points of the image is divided into multiple layers,then local transformation methods is applied within each layer to generate registration candidates,and finally seam-cutting and multi-band blending is performed to fuse the images.The main research contents includes:1.A layer separation method is proposed to generate registration candidates,which greatly reduce the number of registration candidates compare to the randomized methods commonly used by other seam-guided methods,thus the time complexity of the algorithm is reduced.2.Local transformation method is used to register images for each layer using the matching points within that layer.By performing the RANSAC algorithm to filter the feature matching points within each layer,the distortion caused by the local transformation method is eliminated.3.A new energy function for seam optimization is proposed.The quality of the generated seam is improved by introducing gradient error and Lab color space.Several experiments in large parallax scenes is performed to demonstrate the effectiveness of the proposed image stitching method.By comparing with several commonly used algorithms,we show that the proposed method has advantages in both stitching result quality and time performance.