Model-based Local Distortion Flow Estimation for Wide-angle Image Rectification

Wide-angle cameras are important for large-scale surveillance because of the larger field of view. However, due to lens design limitations, it distorts the captured image to enlarge the camera view. The degree of distortion is usually varied according to the position of the object. Nevertheless, a regular and perspective image view is preferred for viewers; thus, image rectification becomes essential. This paper proposed a learning network to estimate the distortion flows coupled with locally-adaptive model fitting to correct the distortion of wide-angle lens images. Unlike some data-driven methods that directly learn the mapping between an input image and its image distortion parameters, we firstly estimated the motion flow between the distorted and rectified images. Next, by fitting a model to locally infer the model parameters, we generated a model-regularized flow map for rectification. Our experimental results show the barrel distortion can be robustly corrected.