为提高单应性估计的准确性和解决真实标注难获取的问题, 提出一种具有修正功能的无监督单应性估计算法。该算法采用级联结构, 其思想类似于迭代, 其中每一级网络都保持相同的层数和参数量, 下一级网络输出的单应性矩阵为真实矩阵与之前输出单应矩阵和的残差。考虑到模型复杂度和实时性的需求, 文章采用两级网络级联。通过在COCO数据集中的5000张图片上进行验证, 结果表明, 相比传统方法和其他基于深度学习的方法, 所设计的级联无监督算法具有更准确的估计能力, 其在测试集中的平均像素误差为0.54, 较传统方法下降95.38%, 运行速度达到95f/s。

In order to improve the accuracy of homography estimation and solve the problem that it is difficult to obtain real labels, an unsupervised homography estimation algorithm with correction function is proposed. The algorithm adopts a cascade structure, and its idea is similar to iteration, in which each level network maintains the same number of layers and parameters, and the homography of the output of the next level network is estimated as the residual of the sum of the real matrix and the previous output homography matrix. Considering the requirements of model complexity and real-time, a two-level network cascade is adopted in this paper. Through the verification on 5000 images in the COCO dataset, it shows that the cascade unsupervised algorithm designed in this paper has more accurate estimation ability than traditional methods and other methods based on deep learning. Its average pixel error in the test set is 0.54, which is 95.38% lower than that of traditional methods, and the running speed reaches 95f/s.