Image Stitching Using Green Coordinates And Improved Seam Estimation

Image stitching is to blend two or more images with overlapping areas into a seamless,high-resolution and wide field-of-view image.The stitched image has a larger field of vision and higher resolution than a single image.Image stitching has gained considerable research interest recently in the field of computer vision and image processing.As the main carrier of human visual information,image has penetrated into all walks of our life.In virtual reality,target recognition,remote sensing technology,medical image processing and other fields,we often need high-resolution image with large perspective.Because of the limitation of viewing angle of ordinary cameras,it is difficult to obtain the required wide field-of-view image,and there fish-eye lens and panoramic cameras in the current market are prone to distortion.Therefore,the further research of image stitching has a very important application value.With the rapid development of computer vision,image stitching is facing great challenges as well as opportunities.In recent years,researchers focus on large-parallax scenes and improving the natural-looking of stitched images.Due to the existence of parallax,the alignment between input images is not as simple as that without parallax,and the situation is much more complicated.There is a certain relationship between the existence of parallax and the occurrence of ghosting in the overlapping region and the appearance of stitched images that do not conform to the human naked eye.Aiming at the problems faced by image stitching,we have made a profound study,the main research includes the following points:(1)In view of the traditional random sampling consistency algorithm RANSAC,which is used to eliminate mismatches,the inaccuracy of the homography matrix as the RANSAC screening model is analyzed in the case of parallax.It is shown that the RANSAC algorithm,which uses the homography matrix as the screening model,is easy to delete the correct matching points and retain the wrong matching points.Since matching feature points play an important role in the process of subsequent image stitching,in order to avoid the bad influence of matching feature points on subsequent image stitching caused by the inaccuracy of matching feature points,we start from the camera imaging principle,find the essential relationship between two images of the same scene in two different perspectives,and then the constraint relationship between the image points of the same three-dimensional point in two images is determined.We then use this constraint relationship,specially,the basic matrix as the screening model of RANSAC,to more accurately delete the wrong matching feature point pairs,and retain the correct matching feature point pairs.(2)Based on Green coordinate and robust elastic warping,propose a new image stitching method.Aiming at the problem of low natural-looking in the final stitched,an unaligned image is warped to the reference image by Green coordinates.And then,the feature points of the pre-warped image and the reference image are matched,and the projection bias caused by the mapping of feature points is calculated,the influence caused by parallax is eliminated by correcting the projection bias.Finally,seam estimation is implemented to seamlessly blend the registered warped and reference images together.The experimental results validated the efficiency of the proposed image stitching algorithm,(3)Aiming at the drawback of traditional seam estimation algorithm,the improved seam estimation method based on superpixel segmentation is studied.The trend of the traditional seam is limited greatly,which is easy to cause the seam to cut significant objects.To solve these problems,the idea of superpixel segmentation is applied to the search of seam,so that the optional seam could pass through the edge of superpixel.The problem of finding the path with minimum energy in the overlapped region of image is transformed into finding the path in the 0-1 matrix of the superpixel boundary,and Depth first search is used to search the path in the 0-1 matrix.