Fisheye Image Calibration and Super-Resolution Method Based On Deep Learning.

Fisheye lens is a kind of ultra-wide-angle lens, which can get more field of view information than ordinary perspective lens. However, there is radial lens distortion in the photos taken by fisheye lens. In order to improve the visual quality of fisheye images and to perform subsequent computer vision tasks, fisheye images must be calibrated first. Traditional calibration methods are often based on a specific type of fisheye lens, which is not universal, and often need to take multiple images of a calibration pattern (typically a checkerboard) for calibration. And limited by hardware, many fisheye cameras take photos with low resolution. The method proposed in this paper not only overcomes the above limitations, but also jointly solves the problem of image super-resolution, which makes the final result more appreciable. This method uses a depth convolution neural network model to predict the fisheye image correction parameters, and then completes the correction process, but also completes the image super-resolution processing. Our quantitative experiments have proved the feasibility and superiority of our method.