Fully Neural Network Mode Based Intra Prediction of Variable Block Size

TL;DR

This paper introduces a neural network-based intra prediction framework for variable block sizes in image coding, achieving better compression efficiency with lower complexity than previous methods.

Contribution

It presents a novel fully neural network mode framework for intra prediction, tailored for different block sizes, and demonstrates improved coding gains over traditional methods.

Findings

Achieves 3.55%, 3.03%, 3.27% BD-rate savings for Y, U, V components.

Uses different neural network structures for small and large blocks.

First to explore a fully neural network mode based intra prediction framework.

Abstract

Intra prediction is an essential component in the image coding. This paper gives an intra prediction framework completely based on neural network modes (NM). Each NM can be regarded as a regression from the neighboring reference blocks to the current coding block. (1) For variable block size, we utilize different network structures. For small blocks 4x4 and 8x8, fully connected networks are used, while for large blocks 16x16 and 32x32, convolutional neural networks are exploited. (2) For each prediction mode, we develop a specific pre-trained network to boost the regression accuracy. When integrating into HEVC test model, we can save 3.55%, 3.03% and 3.27% BD-rate for Y, U, V components compared with the anchor. As far as we know, this is the first work to explore a fully NM based framework for intra prediction, and we reach a better coding gain with a lower complexity compared with the previous work.