Attention-Based Neural Networks for Chroma Intra Prediction in Video Coding

Neural networks can be successfully used to improve several modules of advanced video coding schemes. In particular, compression colour components was shown greatly benefit from usage machine learning models, thanks the design appropriate attention-based architectures that allow prediction exploit specific samples in reference region. However, such tend complex and computationally intense, may difficult deploy a practical pipeline. This work focuses on reducing complexity methodologies, set simplified cost-effective for chroma intra-prediction. A novel size-agnostic multi-model approach is proposed reduce inference process. The resulting architecture still capable outperforming state-of-the-art methods. Moreover, collection simplifications presented this paper, further overhead architecture. Thanks these simplifications, reduction number parameters around 90% achieved with respect original methodologies. Simplifications include framework convolutional operations, cross-component processing model integrated into architecture, methodology perform integer-precision approximations aim obtain fast hardware-aware implementations. schemes are Versatile Video Coding (VVC) pipeline, retaining efficiency intra-prediction methods based neural networks, while offering different directions significantly complexity.