Based on improved unsuperpoint image stitching method
Rui Zhou Zhenzhen Liu, Ming Liang Lu, Xianke He, Zuguo Chen · 2022
When the traditional ORB algorithm is applied in the field of image mosaic, it is susceptible to the interference of light factors and cannot extract high-quality features, which leads to the inaccuracy of the perspective transformation matrix and the problem of poor mosaic effect. In response to this problem, this paper proposes an improved Unsuperpoint feature extraction network to apply to image stitching. By designing the backbone network of the unsupervised model, the points of interest and the descriptor loss function are optimized to improve the efficiency of the network and the accuracy of feature point extraction. Compared with the traditional ORB algorithm, the image stitching algorithm in this paper reduces the image matching time by half and increases the matching accuracy by 7%.The images stitched by this method avoid the phenomenon of cracks and black lines at the joints, and the image transitions are natural and high in definition.