Research On UAV Image Mosaic Technology Based On Grid Deformation

UAV image mosaic is to merge UAV images from different viewpoints and partially overlap into a clear and seamless panoramic image.Unmanned aerial vehicle image stitching requires a series of steps such as feature extraction,feature matching,resampling,uniform color,and overlap fusion to achieve position correction and grayscale correction.However,unlike satellite or ground images,due to the undulation of buildings or the ground,there is parallax between images taken from different positions in the same area,which can lead to significant misalignment and ghosting in the stitching results.The existing stitching methods synthesize images by extracting local feature points,but the biggest drawback is the small number of feature points,which leads to insufficient utilization of image content information and unreliable grid constraints.In response to this issue,this article conducts research and development on the method of eliminating parallax in unmanned aerial vehicle image stitching.(1)In structurally rich natural scenes,captured images have obvious linear structures and corner points,but they are not fully utilized.In order to extract richer feature points,this paper proposes a local elastic distortion method based on line point consistency,and utilizes these features to eliminate parallax.The specific method is to first detect matching points and line segments through SIFT and LSD algorithms,then delimit molecular regions of the image,construct line point invariant,and finally match regions to achieve feature point enhancement.Based on the extracted richer feature points,this paper divides the image into regular grids,estimates the final deformation position of the grid based on thin plate splines,and achieves local registration between images.(2)In scenes with repetitive textures such as farmland,forest land,and bare land,it is difficult to extract effective feature points from drone images,making grid correction impossible at small scales.To address this issue,this paper proposes a drone image stitching method based on deep feature points,which introduces a Transformer network to extract image feature points.Thanks to the self attention mechanism and cross attention mechanism of the Transformer network,more and more uniform matching points can be extracted within the image overlap area.Based on the Transformer’s similarity rating between blocks,this article also designs a new point matching term to guide grid deformation and achieve local alignment of overlapping areas.At the same time,this article also introduces line protection and distortion control to maintain the image structure unchanged.(3)In order to make the above work practical,this article implements a drone image stitching software.This software is developed based on Python and MATLAB,encapsulating the unmanned aerial vehicle image stitching algorithm proposed in this article,achieving automatic image stitching and visual result display.