Rate-Distortion Optimized Learning-Based Image Compression using an Adaptive Hierachical Autoencoder with Conditional Hyperprior

Learning True Rate-Distortion-Optimization for End-To-End Image Compression

“…In [11], randomly generated masks were used during training. These masks were independent of the input image, which does not optimally reflect the use-case, as during inference the mask is determined to minimize the RD loss function.…”

“…With this simple procedure, we are able to efficiently split the image sensibly. Differently to the random initialization from [11], the content which is encoded by each latent space level shares some characteristics, so the latent space can specialize on that characteristic. Preliminary experiments have shown that the precise values of σ 2 th,1 and σ 2 th,2 are not actually critical for the performance of the network.…”

“…After we obtain a trained network, we have to perform a rate-distortion optimization to find the optimal encoding configuration for each image. Here, we first follow the approach from [11]. In this approach, the image is first divided in blocks of size 64×64.…”

“…Therefore, this process does not result to a global optimum. To counter this effect a two-pass RDO was proposed in [11]. The second RDO pass optimizes the blocks again, but this time when all surrounding blocks are initialized with the result of the first RDO pass.…”

“…In this work, we build on RDONet [11], a recently proposed hierarchical autoencoder structure, which enables a dynamic RDO similar to adaptive block partitioning. In this paper, we boost the performance of RDONet with a novel training procedure.…”