改进基于多相机的无监督学习图像拼接算法

Improved unsupervised learning based on multiple cameras image stitching algorithm

首先,针对大范围特征不显著地面图像的获取,由于传统单相机受到视场角和分辨率以及外界环境的影响,提出基于多相机对大范围地面的图像进行采集。其次针对大范围特征不显著地面图像使用传统方法拼接中存在特征提取能力差以及拼接产生的错位、伪影和结构变形问题,提出对无监督学习图像拼接的框架结构改进,提升图像拼接质量。最后借助图像拼接的评价指标对改进前后的拼接图像进行评价。实验结果表明,改进后方法不仅可以有效的解决传统方法基于特征不显著以及纹理相似度高的地面图像拼接中产生的伪影和失真,而且解决了无监督图像拼接过程中产生的结构的变化。改进后的拼接方法拼接后的质量优于传统方法的拼接,并且改进后的算法的迁移性很高,不止针对无显著特征的大范围地面,而且可以广泛用于不同场景下的大范围大基线图像拼接。

Firstly, for the acquisition of large-area feature-insignificant ground images, we propose the acquisition of large-area ground images based on multiple cameras because the traditional single camera is affected by the field of view and resolution as well as the external environment. Secondly, we propose to improve the framework structure of unsupervised learning image stitching to improve the image stitching quality. Finally, the stitched images before and after the improvement are evaluated with the help of the evaluation index of image stitching. The experimental results show that the improved method can not only effectively solve the artifacts and distortions generated in the traditional method based on the ground image stitching with insignificant features and high texture similarity, but also solve the changes of the structure generated in the unsupervised image stitching process. The quality of the improved stitching method in this paper is better than that of the traditional method of stitching, and the improved algorithm is highly migratory and can be widely used for stitching a large range of large baseline images in different scenes, not only for a large range of ground without significant features.