Research On Parallax Image Stitching Based On Point And Line Features

With the increasing use of images and the vigorous development of industries such as artificial intelligence,film and television production,and smartphones,image stitching technology is widely used in video surveillance,unmanned driving,panoramic shooting,medical image registration,surveying and mapping.The large-angle panoramic images synthesized by this technology are gradually appearing in people's daily lives,effectively improving people's living standards and quality.The target scenes for image stitching can be divided into non-parallax scenes and large-parallax scenes.Due to the randomness and complexity of parallax,more advanced equipment and algorithms are only limited to the processing of small parallax scenes,and the image stitching effect of large parallax is still not ideal.The problems of image ghosting,distortion and chromatic aberration caused by parallax and chromatic aberration cannot be effectively overcome.At present,the important image stitching technology still has research space and value.Therefore,this thesis studies the image stitching method for large parallax scenes.In this thesis,the image taken by the handheld device is taken as the experimental object,and the research is focused on improving the alignment accuracy of the image.Based on the traditional splicing framework based on feature points,this thesis proposes some novel and targeted measures.For example,the introduction of line features and image segmentation to improve the number of matches and matching accuracy,the use of local transformation model to map the image,the use of feature constraints for error correction,to improve the performance of image parallax tolerance.The content and innovation of the research in this article are as follows:1.An image matching method based on normal distribution is proposedIn view of the problems of mismatching,insufficient feature matching and long operation time in the current registration algorithm,this thesis proposes an image feature matching measure based on the normal distribution model.First,perform vector operations on the feature descriptors to reduce the dimensions to improve the efficiency of the operation.Then,the introduced line features are converted into point features to enhance the richness of image features and the robustness of image registration.Finally,the superpixel segmentation method is introduced to perform plane segmentation on the image to accurately calculate the local mapping model,and the normal distribution model is used to classify the mapping errors of the point features to identify mismatches,so as to find the mismatch points adaptively.Experimental results show that the method in this thesis effectively reduces the false matching of features,enriches the matching features of images,and improves the efficiency of the algorithm.2.A local alignment model based on bias correction is proposedIn order to solve the problem of misalignment in the local image area of the traditional local transformation model,this thesis proposes a local alignment strategy based on deviation correction.First,the optimized local transformation model is used to project the gridded image.Then,the projection errors of the point and line features are calculated,and a model describing the local deviation is constructed to fit the correction field of the image projection deviation.Finally,the fitted deviation matrix is used to correct errors in the transformed image to improve the alignment accuracy of the image.Experimental results show that the method effectively eliminates image ghosting caused by misalignment.3.A gradual image fusion method based on arc function is proposedIn the actual shooting process,changes in the position of the viewpoint,differences in camera parameters,and changes in light will cause certain parallax and chromatic aberration between images.This thesis proposes a nonlinear weighting model based on circular arc function to perform weighted fusion on the deformed image to solve the unnatural phenomena caused by parallax and chromatic aberration.By calculating the relative position of the current pixel in the overlapping area.Then,use the arc function to calculate the weighting value corresponding to each position.A weighted matrix is used to superimpose the images to fuse the images.Experimental results show that the method in this thesis effectively solves the problems of stitching seam and motion ghosting.