In the last few years, several learning-based image coding solutions have been proposed, mainly using deep neural network models. These solutions claim to advance state-of-the-art in image coding by leveraging recent advances on generative models and other unsupervised image processing problems (e.g. super-resolution, denoising). Typically, these solutions require complex training procedures with large image datasets and exploit the availability of highly parallelizable graphic processing units (GPUs). Recognizing that this area is considered rather promising, JPEG has started studying learning-based image codecs with a precise and well-defined quality evaluation methodology.

In this report, a taxonomy was proposed and solutions from the literature were organized into several classes. The objective is to easily identify and abstract the differences, commonalities and relationships of the learning-based image coding solutions already available. Besides, a list of promising image coding implementations and potential datasets to be used in the future was gathered.

To keep informed and get involved in the discussion, interested parties are invited to register to the ad hoc group’s mailing list.