Model-free prediction of atmospheric warp based on artificial neural network

Abstract

This paper presents the application of artificial neural network for predicting the warping of images of remote objects or scenes ahead of time. The algorithm is based on estimating the pattern of warping of previously captured short-exposure frames through a generalized regression neural network (GRNN) and then predicting the warping of the upcoming frame. A high-accuracy optical flow technique is employed to estimate the dense motion fields of the captured frames, which are considered as training data for the GRNN. The proposed approach is independent of the pixel-oscillatory model unlike the state-of-the-art Kalman filter (KF) approach. Simulation experiments on synthetic and real-world turbulence degraded videos show that the proposed GRNN-based approach performs better than the KF approach in atmospheric warp prediction.

Blind atmospheric turbulence deconvolution

“…A two‐steps method using first a multiscale optical flow estimation and then the first register then average and subtract algorithm was proposed in [15, 16] to obtain a restored image. The authors in [17, 18], respectively, use a generalised regression neural network and a convolutional neural network to learn turbulence‐induced deformations. These neural networks are then used to predict and compensate for the turbulence impact.…”