基于生成对抗网络的极坐标域自监督径向畸变校正算法

Self-calibration radial distortion correction algorithm in polar coordinate domain based on generative adversarial networks

鱼眼相机可以捕捉大视野场景,但它在图像中引入了严重的径向畸变.大多数现有的深度学习图像校正方法都需要畸变参数、无畸变图像等标签进行监督,而应用到真实世界的场景时,上述的假设和要求可能并不成立.提出基于生成对抗网络的极坐标域自监督径向畸变校正算法,利用极坐标图像畸变大小水平一致性和极坐标图像内部畸变平滑性,实现单幅图像的自监督鱼眼校正.同时使用多项式模型和除法模型合成大范围的畸变训练数据集,使模型获得较好的泛化能力.在合成数据集和真实数据集上的对比实验结果证明了本方法的优越性.

Fisheye cameras can capture scenes with a large field of view,but they introduce severe radial distortion in the image.Most existing deep learning-based methods require corresponding calibration labels such as distortion parameters and distortion-free images,etc.,while the above assumptions and requirements may not hold when applied to real-world scenarios.Therefore,this paper proposes a selfcalibration radial distortion correction algorithm based on generative adversarial networks in the polar coordinate domain.This is to achieve self-calibration fisheye correction of a single image by exploiting the horizontal consistency of distortion size in polar coordinate images and the smoothness of distortion within polar coordinate images.The polynomial model and the division model are used to synthesize a large range of distorted training datasets so that the proposed model obtains a better generalization capability.The comparison experimental results on the synthetic and real datasets demonstrate the superiority of the method.