Tiki-taka Training based End-to-end Variational Auto-encoder Image Compression

Recently, learned image compression has shown remarkable progress, and its compression efficiency has outperformed the traditional image coding standard like JPEG2000. However, existing learned methods lack the mechanism to extract the adaptive features well or lack the exploration of the quantization asymmetry between the training and inference phase. In this paper, a novel learned image compression framework is proposed. Firstly, the asymmetric convolution block (ACB) is integrated into the variational auto-encoder framework, which can increase the feature extraction capabilities of the convolutional neural network (CNN) by enlarging the weight of important locations. Secondly, the global context from long short-term memory (LSTM) and local context from mask convolution are combined to estimate the probability density model parameters. Finally, a tikitaka training strategy is proposed to solve the asymmetric error caused by the quantizer between the training and inference phase. Experimental results show that the proposed scheme provides a better rate-distortion performance, which can achieve 6.35% BD-rate saving than the baseline on average.