Image De-Raining for Driver Assistance Systems using U-Net based GAN

The visual appeal of images gets deteriorated when captured under unfavorable weather conditions thus rendering them unsuitable for optimum performance on vision-based systems. Performance of Driver Assistance Systems (DAS) depends on the quality of the images fed to the computer vision algorithms. This paper attempts to leverage generative modelling capabilities of Generative Adversarial Networks by utilizing special architectures for generator and discriminator. The loss function utilized is the efficient pixel mean square loss. Discriminator used here is a standard binary cross-entropy classifier intended to classify whether the predicted de-rained image of the generator matches the real high-resolution image or not. The generator network is a Fully Convolutional U-Net architecture. Experiments indicate that the proposed architecture produces de-rained images indistinguishable from their clean images using only 30 epochs and performs better than various novel approaches.