Large-scale Image Stitching Algorithm Integrating UAV Position Information and MGFT Model
Zun Liu, Junxin Liu, Kaiwen Wang, Jinxiao Tan, Jianqiang Li · 2024
We propose a large-scale image stitching algorithm integrating UAV position information and MGFT(Max Grad Feature with Transformer) model to deal with the issues of slow speed, low accuracy, and projection distortion in large-scale image stitching. First, the algorithm calculates the image splicing sequence through the special barrel structure of the containment vessel and the UAV’s position information, reducing the search space for large-scale image splicing and reducing ineffective image matching, thus greatly improving the efficiency of image splicing. Secondly, the algorithm implements a method of estimating the transformation matrix parameters of three types of images through UAV position information, and proposes a method for direct splicing using UAV position information. Finally, to estimate the translation matrix parameters more accurately, we propose the MGFT model. By stacking multiple layers of Transformer’s self-attention and cross-attention, the extracted maximum pixel gradient features can integrate the global information of two images. Experimental results show that the algorithm proposed in this article improves the efficiency of large-scale image stitching, improves the accuracy and quality of image stitching, and reduces cumulative errors and projection distortion.