Spatio-Temporal Fish-Eye Image Processing Based on Neural Network

In this paper, we describe an approach to correct the distortion in fish-eye image. Due to pros of large FOV (field of view), fish-eye camera is widely used in computer vision discipline. However, the serious distortion in fish-eye image set up barriers to image processing. This paper focus on adapting neural network for correcting distortions in fish-eye image. Up to now, traditional correction models such as latitude-longitude, sphere and grid template establish a certain model that is ideal and may not fit the reality situation. The method used in our study is known as neural network. Instead of approximately regarding the fish-eye distortion model as high-order polynomial function, we use neural network to learn the relationship between the distorted and corrected image. The mean square error for the difference between the corrected and ideal points is 4.1345 pixel per point, which is much smaller than the result of polynomial model (32.0809 pixel per point), and the run time for this algorithm is moderate. The results of the experiment indicate that correction model based on neural network can solve the fish-eye distortion in an acceptable error and high efficiency.