Rate-Distortion Optimized Learning-Based Image Compression using an Adaptive Hierachical Autoencoder with Conditional Hyperprior
Fabian Brand, Kristian Fischer, André Kaup · 2021
Deep-learning-based compressive autoencoders consist of a single non-linear function mapping the image to a latent space which is quantized and transmitted. Afterwards, a second non-linear function transforms the received latent space back to a reconstructed image. This method achieves superior quality than many traditional image coders, which is due to a non-linear generalization of linear transforms used in traditional coders. However, modern image and video coder achieve large coding gains by applying rate-distortion optimization on dynamic block-partitioning. In this paper, we present RDONet, a novel approach to achieve similar effects in compression with full image autoencoders by using different hierarchical levels, which are transmitted adaptively after performing an external rate-distortion optimization. Using our model, we are able to save up to 20% rate over comparable non-hierarchical models while maintaining the same quality.