Research On Image Mosaic Algorithm And Anti-aliasing Mechanism Based On Thin Plate Spline

Image stitching refers to the fusion of two or more images with a narrower field of view into a panoramic image with a high resolution or a wider field of view through image processing technology.Image stitching mainly includes two stages of image registration and image fusion.In this technology,the stitching of images with parallax has always been a hot issue for researchers.When stitching images with parallax,it is difficult to find a model describing the relationship between the images to be stitching,so it is necessary to remove the parallax from the images.In 2018,Elastic Local Alignment(ELA)algorithm used the thin plate spline Alignment model to eliminate the parallax of the image to be stitched,and then used homography transformation to splice the image to be stitched,which ensured that the image to be spliced had a high enough registration accuracy in the registration stage.It also provides a good idea of parallax image stitching for this paper.This also provides a good idea of image stitching for this article.But at the same time,the ELA algorithm still has the problem that the parallax cannot be completely eliminated,and the final stitched image has the phenomenon of misalignment.In the non-linear transformation of the thin-plate spline alignment model to eliminate the parallax of the original image,if the original image is not pre-processed for anti-aliasing,it will affect the final stitched image quality.The traditional standard Gaussian anti-aliasing operation cannot effectively deal with the aliasing generated by the original image in the nonlinear transformation,so that the image generated by resampling still has local distortion,artifacts,etc.,which affects the subsequent image processing results,including the accuracy of feature extraction.Meanwhile,it is also noted that universal algorithms(including ELA algorithm)consume extra time to extract and match the features of the whole input image by using SIFT(scaleinvariant feature transform)feature to obtain the transformation model in the image registration stage.Based on the above-mentioned problems,this article starts with the stitching speed and stitched image quality of stitched images,and has carried out the following work:(1)Based on the speed of image stitching,an optimization algorithm for image stitching based on the SURF(speeded up robust features)feature to locate the overlapping area of the image is proposed.In the image stitching technology,the stitching image only partially overlaps,so only the feature point pairs in the overlapping area need to be calculated for the transformation relationship General image stitching technology uses SIFT feature extraction from the full image.Considering the slow calculation of SIFT features,this paper proposes to use the faster and better SURF feature points to locate the overlapping area.Then re-extract the SIFT features from the overlapping area to obtain the transformation relationship,thereby improving the speed of image stitching.And the effectiveness of the algorithm is verified through experiments.(2)Based on stitched image quality,a thin-plate spline transformation anti-aliasing method based on mapping adaptive convolution is proposed.The thin-plate spline transformation does not satisfy the bijective definition under certain conditions,so it is not fully applicable to the application conditions of the mapping adaptive convolution.This paper proposes a local affine approximation method based on the analysis of the deformed convolution kernel of the mapping adaptive convolution.And through mathematical analysis and experimental proof to verify the mapping adaptability and anti-aliasing properties of the thin plate spline on the mapping adaptation to the convolution,and lay the foundation for the next image stitching algorithm.(3)On the basis of(2)and based on the stitched image quality,propose a two-sided thinplate spline image stitching algorithm based on mapping adaptive convolution.This paper sorts out the process of ELA algorithm that can tolerate large parallax,and is inspired by the idea of adjusting parallax through thin plate splines.Based on the theory of mapping adaptive convolution,a two-sided thin-plate spline image stitching algorithm based on mapping adaptive convolution is proposed The motivation of this algorithm is that the stitching results generated by the ELA algorithm still have local misalignment,so the second thin-plate spline transformation is used to fine-tune the resampled image,so that the stitched image has a better alignment effect in the overlapping area.It also explains how to obtain the second thin plate spline based on(2),and finally verifies the effectiveness of the algorithm through experiments.