Implementation Of Image Stitching System Based On Feature Matching

Image stitching technology is to synthesize the captured image sequences into a large image,so that it can well meet the needs of people for wide-field images in daily life,work and learning.The essence of image stitching is to use the spatial transformation relationship between the images to be stitched to register,and then maps the images to the same coordinate system to form a wide field of view image.The core of image stitching technology is image registration.Image registration methods can be roughly divided into three classes: registration using gray information,registration using image transformation domains,registration using image feature points.Registration methods based on feature matching have the following advantages: 1.The extraction process of feature points can reduce the influence of noise,and has good robustness to image occlusion,illumination change and so on.2.The number of feature points is much less than image pixels.Using feature points to match can reduce the computational complexity in the registration process.3.The matching metric of the feature points is less affected by distortion,which can improve the accuracy of registration.Therefore,this paper will take advantage of the registration methods based on image features points to achieve image registration.SIFT(Scale-Invariant Feature Transform)feature has good properties in many image features.It is not only insensitive to direction,but also robust to rotation,scale and other transformations.However,because of the large redundancy of the key points extracted by SIFT algorithm and the random distribution of feature points,the computation is heavy and the operation efficiency is low.The SURF(Speeded-up Robust Features)algorithm improves the detection and description of SIFT feature.It makes the operation more efficient and faster by using integral images and box filters.So,this paper will utilize SURF feature to realize image registration.Only the spatial transformation relationship between two images is considered in the image registration process,the global information between the images to be stitched is not considered.If the registered images are directly mapped to the same coordinate system,the waveform phenomenon of the stitched wide-field image will occur.Therefore,it is necessary to correct the registered images globally so that they are on the same horizontal line as much as possible.In addition,if the registered images are mapped directly,the obtained wide-field image may have large distortion on both sides.In order to eliminate these distortions,it is necessary to establish a mapping model for images to be stitched.In this paper,three mapping models will be discussed.That is,plane mapping,cylinder mapping,spherical mapping.After that,a high-quality wide-field image can be obtained through stitching seam search,illumination compensation,image fusion and other steps.In this paper,the stitching of image sequences with rotational motion,elevation change,depression angle change,scale change and so on is realized,and a wide-field image is obtained.We can correct the objects in the wide-field image to keep the real state as far as possible.In addition,we also can effectively eliminate the distortion around the wide-filed image,and the final wide-field image is more consistent with human vision.