Research & Implementation Of Panorama Stitching Based On Feature Matching

In recent years,with the rapid development of intelligent devices,terminal devices,autonomous driving,robot computing,and Internet technology,many applications propose strict requirements for image perception.Based on this background,this paper aims to study high-quality image stitching algorithm,and carry out engineering development and implementation.Firstly,the paper introduces the basic background,research significance and research content of image stitching.Then,based on the basic principles of image transformation,we describes various types of image mapping transformation,including rigid body transformation,affine transformation,projection transformation and non-linear transformation,etc.Then the image stitching algorithm based on feature matching and homography matrix regression algorithm based on deep learning are presented.Finally,the image stitching service is developed and deployed in AWS cloud,and the API calling service is realized.The work and innovation of this paper could be list as follows:The SURF description operator is detailed to extract image features,and the FLANN algorithm is exploited to match the image features.In order to improve matching quality,we apply RANSAC algorithm to filter out outliers and then update the homography matrix using inliers,and finally the homography transformation is applied to achieve image stitching.Based on the limited image data to be stitched,the method of constructing the true value of image group and homography matrix is proposed,which helps us to construct thousands of image matching group data sets and their ground truths(homography matrix parameters);then a convolution neural network is designed and developed based on simulation data,and the regression target of neural network is 8 homography matrix parameters;which is then implemented based machine learning framework TENSORFLOW to achieve the image homography matrix regression.After the homography matrix of the image pairs is obtained,we use the gradual fade out method to post-process of the projected image.The fusion results are clipped then to get the final image,which can quickly solve the problem of different exposure levels of image fusion.At the same time,we also developed the development stack of cloud deployment based on AWS services,using API gateway,lambda function and ECR image.We can invoke the image stitching service based on HTTP request.The request adopts post method,carries base64 encoding of input image pairs,and returns the fused image data with JSON format.