Research And Implementation Of Large Parallax Image Stitching

Image stitching technology is designed to join two or more images taken at different viewing locations into one seamless wide-view panorama through aligning and superimposing overlapping areas of these images.By synthesizing the images obtained from multiple shots,this technology enlarges the view range of the photo and makes the photo content no longer limited to the range that the lens can capture.Nowadays,this technology has been widely used in many different fields such as image recognition,video surveillance,computer vision and so on,which has important research significance and practical value.However,when the photographed 3D scenes are not in one plane and the baselines of the shots are very different,the relative positions of the scene structures in the resulting images are completely different.Existing image stitching algorithms cannot be used to stitch such images with large parallax.Based on the above considerations,this paper focuses on the image stitching problem with large parallax images,and the specific situation that the two to be stitched images respectively contain the left-side and right-side faces of a human is considered.Specifically,the main works of this paper are as follows.(1)Based on the 3D rotation modeling to the key-points in left and right faces,the reconstruction of key-points in the frontal face is realized.Firstly,the key-points of the two faces in the two images are extracted,and a 3D rotation model on the face key-points is constructed by studying the camera imaging model.Secondly,based on this model,the rotation relationship between the key-points in left and right faces is summarized as a set of unconstrained optimization problems.By replacing the obtained values with parameters in the model,the rotation path of the face key-points from left to right can be obtained,as well as the coordinates of the key point of the face at any rotation angle during this rotation.Finally,by setting the rotation angle to half of the whole rotation angle,all the coordinates of the frontal face key-points are obtained.(2)The large parallax image stitching is achieved by separately stitching and fusing the face region and the background regions.Firstly,based on the frontal face key-points,the frontal face image is obtained by stitching the face region using triangular dissection.Secondly,the transformation relationship between the key-points is applied to the nonface regions,by twice homograph transformation and the proposed center point radiation distortion algorithm,the non-face region is corrected,and the background region stitching is then achieved by properly adjusting the image opacity.Finally,the large parallax stitching image is obtained through seamlessly stitching the frontal face and non-face region by Poisson fusion algorithm.Experiments show that the proposed algorithm can accurately achieve large parallax image stitching efficiently.(3)An online platform for image stitching is built based on the proposed largeparallax image stitching algorithm in this paper.The platform mainly consists of four functional modules,and the main function is the large-parallax image stitching.The platform promotes the practicality of image stitching and effectively improves the image recognition accuracy.