Image Stitching via Augmentation Selection and Comprehensive Optimization of Homography Models
Wanli Xue, Yuanyuan Chen, Yanhong Yang, Weilun Xie, Shengyong Chen · IEEE Transactions on Consumer Electronics · 2023
The naturalness of prominent objects is important information in image stitching research, which directly affects the satisfaction of electronic consumers when acquiring large scenes. Most current image stitching algorithms use the best homography model for registration features, which usually cannot guarantee the integrity of static and dynamic objects. Aiming to alleviate the problem, we propose an image stitching method in this paper. Specifically, it consists of two parts, augmentation selection and comprehensive optimization of homography models. We first obtain multiple homography models and further refine the model results by cluster analysis to form a reasonable homography candidate set, next the most reasonable homography model and best stitching result can be obtained with the maximal evaluation value. Then, mesh optimization and multiple matching are performed on each model to eliminate artifacts, while optimal seams are optimized to reduce distortion. Extensive experimental results show that our method improves the eval scores by 10.15%, and reaches superior performance over integrity, primitiveness, and uniqueness on both static and dynamic markers.