A Deep Learning Approach for Rectifying Fisheye Image

Fisheye lens cameras offer a significantly wider field of view compared to conventional cameras. However, they also introduce geometric distortions and distorted shapes in captured images. This paper presents a self-supervised fisheye correction method based on convolutional neural networks. The method is trained using a readily available urban landscape roadway dataset and leverages the inherent stability of pixel coordinate relationships after uniform distortion to design an effective correction model. During the training process, distorted images are synthesized from the original undistorted images, and the predicted parameters are utilized to correct fisheye images. Furthermore, the loss function is constructed using the geometric mean of the Hough transform, and the entire model is trained with the assistance of the slope loss of the straight line, thus achieving an end-to-end framework. Experimental results demonstrate that our method provides superior performance in correcting real-world in-vehicle images, outperforming similar approaches in terms of correction accuracy.