Study And Implementation Of Image Stitching Under Near-object Distance And Large Parallax

Image stitching has long been an important issue in computer graph and vison,which is widely applied in the panoramic video surveillance,remote sensing map,virtual reality etc.Various methods on image stitching has been proposed these years but none of them can handle with large parallax well,which makes study on parallax-tolerant image stitching algorithm very significant.Aimed at handling the large parallax during the stitching on images acquired in BaoSteel's warehouse,the global and local registration are combined in this paper.Multi-feature detection is employed to improve the efficiency of global alignment and the region constraint is added during the local refinement to raise the accuracy.Moreover,a new way called “wholerow steel file” is used to archive the steels' images.And the hardware system and software system are designed and realized according to the real scene.The main achievement is as following:(1)Aimed at the problem of large disparity between adjacent images,the conventional method using only single aligning model is improved.A loose global homography is used to match the images roughly and the local alignment based on mesh optimization is applied to refine the local regions later,in which way the ghosting and mis-matching can be reduced efficiently.(2)A fast global image alignment model using multi-feature detection is proposed.Before SIFT feature's detecting and matching,the high speed binary operator FREAK is used to confirm the overlapping areas between images firstly to reduce the scope for SIFT.Also,an improved RANSAC algorithm is applied to raise the accuracy.(3)The conventional mesh optimization is combined with mutual information to preserving both geometry and region grey information.Considering the large parallax,a series of energy constrains functions are designed,including registration constraint,geometric constraint,similarity constraint,linear constraint and the maximal mutual constraint.(4)A new way called “steel file” is used to archive the steels' images in which the global images of the whole row are store and also the high definite image of every single steel is stored as well.The hardware and software system of panoramic image stitching is designed and implemented,in which the image's acquisition,transmission,stitching and display are fully automatic.By combining and improving the conventional algorithms,the algorithm proposed in this paper can handle large parallax well during the image stitching.The experimental results show that,compared with the classical AutoStitch,the recent APAP and SPHP,the proposed algorithm can avoid ghosting and mis-matching more efficiently.