Image coding algorithms create compact representations of an image by exploiting spatial redundancies and perceptual irrelevance, thus exploiting the characteristics of the human visual system. Recently, data driven algorithms such as neural networks have attracted a lot of attention and have become a popular area of research and development. This interest is driven by several factors, such as recent advances in processing power (cheap and powerful hardware), the availability of large data sets (big data) and several algorithmic and architectural advances (e.g. generative adversarial networks).

Nowadays, machine learning through neural networks is the state-of-the-art for several computer vision tasks, such as those requiring high-level understanding of content semantics, e.g. image classification, object segmentation, saliency detection, but also low-level image processing tasks, such as image denoising, inpainting and super-resolution. These advances have led to an increased interest in applying deep neural networks to image coding, which is the main focus of the JPEG AI activity within the JPEG standardization committee. The aim of these novel image coding solutions is to design a compact image representation model that has been obtained (learned) from a large amount of visual data and can efficiently represent the wide variety of visual content that is available today. Some of the early learning-based image coding solutions already show encouraging results in terms of rate-distortion (RD) performance, notably in comparison to conventional standard image coding (e.g. JPEG 2000 and HEVC Intra) which compress images with hand- crafted transforms, entropy coding and quantization schemes.

This proposal for a PCS 2021 Special Session on Learning-based Image Coding gathers technical contributions that demonstrate the efficient coding of image content based on a learning-based approach. This topic has received many contributions in recent years and is considered critical to the future of image coding, especially the solutions for which learning-based tools substitute the previous conventional architectures, adopting end-to-end training. This special session proposal collects a wide range of contributions on this topic, namely, non-linear data transformations, probability models for entropy coding, block-based coding structures, rate-allocation procedures, intra prediction tools, prediction filters and complexity optimizations. Moreover, a recent subjective quality study performed in the context of JPEG AI is also included, where relevant learning-based image coding solutions are assessed, and the potential of this novel coding approach is analyzed.