Block-Wise Intra-Prediction of Imaging Data Based on Overfitted Neural Networks with On-Line Learning

Abstract

Block-wise intra-prediction is a key technique used by modern video codecs to reduce the amount of data to be compressed. Recently, machine learning (ML) has successfully improved block-wise intra-prediction by employing neural networks. Notwithstanding, the performance of such ML-based methods depends on the amount, quality, and relevance of the training data. Furthermore, they require signalling the learned parameters into the bitstream to be able to reconstruct the original data after decompression, thus increasing bitrates. This work proposes a novel block-wise intra-prediction strategy based on fully connected neural networks (FC-NNs) that avoids the two aforementioned shortcomings within the context of lossless compression. To do so, shallow FC-NNs are used, whose parameters are refined in an on-line manner using only the data being predicted. This allows to accurately fit the FC-NNs to the data of interest and replicate the optimization process, avoiding signaling the learned parameters. Experimental results indicate that the proposed ML-based intra-prediction strategy can outperform the intra-prediction used by modern video codecs with prediction accuracy gains of up to 7.01 dB PSNR.