Radial distortion correction method based on depth separable network

Fisheye camera has been widely used in our daily life with its wide field of vision, and at the same time it also has serious distortion. Radial distortion is the most common form of distortion, and Zhang Zhengyou calibration correction rule is the most common radial distortion correction method. However, it has some problems, such as complicated calibration process and easily influenced by calibration target quality. Therefore, we propose a distortion parameter prediction model based on depth separable network, and directly establish the mapping relationship between fisheye image and distortion parameter to solve the above problems. Firstly, the loss function and learning rate of the model are determined through a series of experiments, and compared with Zhang Zhengyou calibration method and double longitude method in three index values SSIM, PSNR, NRMSE and subjective vision. The results show that our proposed method has the best performance.