Chroma Intra Prediction with attention-based CNN architectures

Neural networks can be used in video coding to improve chroma intra-prediction. In particular, the usage of fully-connected networks has enabled a better characterisation of the cross-component information, improving the prediction accuracy of related linear models. Nonetheless, such multi-layer perceptron architectures tend to disregard the locations of individual reference samples when predicting the chroma block components. This paper proposes a twofold network architecture based on a novel attention module which takes into account the spatial location of each neighbouring reference sample when computing the prediction. The proposed approach is integrated into the Versatile Video Coding (VVC) prediction pipeline. Experimental results demonstrate compression gains over the latest VVC anchor, as well as over previous chroma intra-prediction methods which are based on neural networks.