PDB Unet: A spatio temporal video Fixed Pattern Noise removal network

In this paper we propose a novel video non uniformity correction algorithm based on a convolutional neural network. Fixed pattern noise (FPN) is a temporally coherent noise present on videos due to the non-uniformities of the sensors that can exhibit spatial correlation. This is a common problem with infrared video, degrading image quality and hampering subsequent applications. FPN removal has received less attention than other video restoration problems, and until very recently existing neural network approaches were limited to single frame processing. In this work we present a novel network architecture that takes several frames as input and outputs the estimated FPN. We also introduce parallel vertical & horizontal downsampling branches in the network that amplify the receptive field and help capture better the spatial correlation of the signal. We demonstrate the effectiveness of our method with extensive experiments on synthetic FPN comprising white, row and column Gaussian noise. Quantitative and qualitative comparisons against previous methods show that the proposed architecture can better leverage the spatial and the temporal information to remove the FPN, leading to state-of-the-art results.