Research On Panoramic Image Stitching Based On Improved Registration Algorithm

In recent years,panoramic images are widely applied in fields like medical science,surveying and mapping,virtual reality,computer vision,etc.Applications based on panoramic images also crop up here and there,such as panoramic video monitoring,Google street view,etc.However,the camera which has direct access to panoramic images is usually complicated to operate.It requires professional skills and is at a high price.These factors make it hard to be put into actual use.The most common method is to collect several coincidence scenes and put them together into a panoramic image with a broad perspective,high definition and strong sense of reality by image registration and fusion.The present research discusses the technology of image mosaic of panoramic images by conducting experiments on two different kinds of images,namely,images taken by a fixed rolling camera and images taken by a hand-held camera.Since images with coincidence scenes taken by a fixed rolling camera only have a single rotation transformation,a model of global image transformation can be applied to image registration.In this paper,the stable feature points in images to be spliced were extracted based on the classic SIFT algorithm.Questions were evaluated according to a model of global image transformation,and an improved Random Sample Consensus(RANSAC)algorithm based on the Determinantal Point Processes(DPP)was presented in this paper.This approach used the global negative correlation characteristic of the Determinantal Point Processes sampling method to model the feature points extracted by SIFT,realized the homogenization and decentralization of the sampling points,and eliminated more mismatching points.Then the point set extracted by the Determinant Point Processes was used as the input of the RANSAC algorithm to obtain the global image transformation model.Experiments proved that the improved RANSAC algorithm based on the determinant point process could ensure that the correct matching rate is over 95%,thus reduced the iteration number of the algorithm,and guaranteed the accuracy of the optimal global image transformation model.The splicing experiments of panoramic images showed that the improved algorithm could well adapted to cylindrical panoramic image mosaic,and the final panoramic image without significant distortion.Since the images taken by a fixed rolling camera has simple image transformation,there was no parallax problem and excellent registration results could be easily achieved by using the global image transformation model.However,when the experimenter held the camera to take required images,It could not guarantee the camera is completely fixed,so great parallax between images took place,and as a result,the global image transformation model could not be applied any more.To solve this problem,the As-projective-as-possible warp(APAP)image registration algorithm based on multi-homography matrix feature detection algorithm was presented in this paper.Multi-homography matrix feature detection was used to detect the same feature point pairs,which avoids the false exclusion of feature points pairs of the parallax images in different planes,and ensures the number of correct feature point pairs.The improved APAP image registration algorithm utilizes the image segmentation technique to divide the image into several small pieces,calculates the locally weighted projection transformation model based on the Gaussian distance between the central pixel and the matching feature point of these pieces,and ultimately uses this model to complete the registration of the coincident area.The experiments showed that the improved APAP image registration algorithm can effectively eliminate the problem of blurred ghost caused by parallax and can synthesize cylindrical panoramic images more clearly.