Artifacts reduction for very low bitrate image compression with generative adversarial networks

Image compression at very low bitrate has known a new dawn since the democratization of convolutional neural networks (CNNs). A promising approach consists in using an « end-to-end » neural network-based compression scheme, performing both coding and decoding phases. Such techniques lead to perceptually convincing results, notably with GAN-based architectures. However, due to the nature of GAN, alongside being computationally expensive on the encoder side, such schemes cannot ensure the preservation of fine details, such as small digits or letters on an ID card. To overcome this limitation, we propose a new model, specifically trained to reduce artifacts resulting from strong lossy compressions. An advantage of the proposed approach comes from the fact that it can be interpreted as a post-processing step that can easily be added to any compression scheme without modifying the codec. Experimental results show that our model perceptually outperforms the state-of-the-art compression standards for very low bitrates.