| Image stitching is the process of stitching multiple images into one panoramic image with a wider field of view.There are usually obvious geometric structures in the scene,such as straight lines,circles,ellipses,or smooth curves.If these edge structures are unreasonably deformed during the image stitching process,the naturalness of the stitching result image will be affected to a certain extent,and even the image content will be obviously distorted.However,most existing image stitching methods ignore these edge information,which can reflect the local details and overall structure of the image,resulting in poor stitching quality.In order to solve this problem,this paper proposes an image stitching method with geometric structure protection,which protects the edge structure in the image to be stitched during the image stitching process,so as to obtain a more realistic and natural image stitching result.The main research contents are as follows:(1)Large-scale edge extraction methods in images.First,the traditional image line detection method and the convolutional neural network-based edge detection method are used to extract important edge structures in the image to be stitched.Then,the extracted edge structure data is further post-processed,such as skeletonization,branch clipping,broken line reconnection and collinear constraints,etc.,to improve the integrity of the structure.Finally,large-scale edge information reflecting the geometric structure of the scene is obtained.(2)A geometric structure preservation method based on triangle sampling.First,triangular sampling is performed on the obtained large-scale edge structure to obtain multiple sets of sampling triangles that can constrain their corresponding geometric structures.Then,the similarity transformation constraint is performed on the sampled triangles in the image deformation to protect the key geometric structures in the image from severe distortion.Finally,a large-scale geometric structure protection model is obtained.(3)A GSP-based geometric preserving image stitching model.Based on the GSP(Global Similarity Prior)image stitching model,the proposed geometric structure protection item is integrated into the model as a new optimization item,and the proposed GES-GSP(Geometric Structure Preserving Warp based GSP)stitching model is obtained.The core idea of the model is to adopt a geometric structure constraint weight adaptive generation strategy,which achieves an effective balance between alignment and geometric structure protection in overlapping regions of images,so that the entire image can be smoothly deformed to present a more natural panoramic effect.(4)Experimentation and analysis of the proposed GES-GSP image stitching method.A diverse and challenging image stitching dataset is constructed,and the proposed method is compared with the stitching results of several representative image stitching methods on this dataset.The proposed method is validated by in-depth analysis from the aspects of geometric structure sampling strategy,geometric structure sampling point constraint weight strategy,subjective evaluation guided by mesh deformation,quantitative evaluation of image quality,and quantitative evaluation of image content distortion to verify its robustness and effectiveness.Experimental results show that GES-GSP outperforms representative methods in terms of image content preservation,and the synthesized panoramic images have more natural visual effects.(5)Design and implementation of mobile image stitching system based on GES-GSP.In order to facilitate the acquisition of natural panoramic images with less distortion and good details,a mobile image stitching software based on the GES-GSP model was developed.The software is designed based on C/S architecture,optimizes and accelerates GES-GSP,and supports users to stitch the images easily.In addition,the software is easy to use and has a friendly operation interface,which provides a new way for users to obtain panoramic images. |
PDF Full Text Request