Research On Image Redundancy Filtering For Panorama Stitching

With the rapid development of UAV remote sensing and sensor technology,it is more convenient to acquire high-resolution ground images.Limited by the flying height of the UAV and the focal length of the camera,it is difficult to obtain the information of the whole target area in a single image.If the focal length of camera is adjusted and the wide-angle lens is used to increase the field of vision,the image resolution will decline and the image edge will be distorted.Therefore,to obtain high-quality large-view panorama image,it has to use panoramic stitching technology to stitch several small-view high-quality images with overlapping areas.The quality of stitching results directly determines the following analysis work,so it is necessary to develop panoramic stitching technology.In recent years,panoramic stitching technology has developed rapidly,which expands from the remote sensing to virtual reality,medical image analysis and so on.In this paper,the problem of image redundancy and global alignment in large-view panoramic stitching task is studied.For large-view image panorama stitching,with the increase of the number of pre-stitched images,the computational cost will be geometrically increased,which not only has a higher demand on the computer performance,but also greatly reduces the stitching efficiency.To avoid missing information of target area when collecting ground images,camera often photographs the target area at a high frequency,which results in some images have a high area of overlap with their adjacent images.If those images are removed from original image set,which will not cause loss information of the stitching results,and will not break the connectivity of filtered image set,those images are called information redundancy.In large-view panorama stitching task,image redundancy will reduce the stitching efficiency and have a higher requirement for computing performance.To solve above problem,under the circumstance of information integrity of stitching results and connectivity of filtered image set,this paper proposed to filter the redundancy image by similarity analysis,first,the similarity matrix of the original image set is obtained by feature points matching,then,based on the similarity matrix,the redundancy images is filtered from original image set by analyzing the similarity relationship in an iterative algorithm.In alignment process,the parallax will inevitably cause the error accumulation,and different alignment model will cause different types of error accumulation,the alignment model with high degree of freedom can get higher alignment precision but can lead to perspective distortion accumulation,the alignment model with low degree of freedom can well resist perspective distortion accumulation but get low alignment precision,which cause local area to appear misalignment.With the increase of the number of images,the above two phenomena will become particularly obvious.To reduce the error accumulation,improve the alignment precision while reducing the perspective distortion accumulation,first,this paper find the optimal reference image and alignment path,then proposes a transition alignment model combining group similarity model and global homography optimization.Firstly,based on the weighted topology graph of the image set,the multi-source shortest path algorithm is used to find the optimal reference image which near the center of the target region,and the shortest path from each image to the reference image is taken as the optimal alignment path.Then,the alignment path is constructed into a hierarchy tree,in which the reference image is taken to group image set.Finally,the group similarity model and global homography optimization were combined to align the images.The group similarity model performed rough alignment to resist the perspective distortion accumulation.Under the constraint of the group similarity,the global homography optimization is employed to improve the alignment accuracy,which eliminates the misalignment in the local area.To prove the effectiveness of image redundancy filtering method and the superiority of the proposed transition alignment model,this paper selects the representative image sets to test.The experimental results show that the image redundancy filtering algorithm can effectively remove the redundant images from the original image set,which dose not cause information loss for stitching result and does not destroy connection of filtered image set and.The running time comparative experiment proved that it can greatly improve the stitching efficiency.The qualitative and quantitative analysis of the stitching results proved that the proposed alignment model can well improve the alignment accuracy while reducing the perspective distortion accumulation.