Abstract
Researchers from Deep Render Ltd have developed a high-efficiency image compression system that combines a nested hierarchical hyperprior with a streamlined context model. By swapping standard activation functions for Padé Activation Units (PAU) and optimizing the PixelCNN++ architecture for a 10x speedup, they achieved state-of-the-art visual quality at low bitrates for the CLIC 2020 challenge.
Introduction
Traditional image codecs like JPEG use hand-crafted algorithms to remove redundancy. While neural network-based "learned" compression has recently surpassed these standards, two major hurdles remain: computational complexity and visual artifacts. Autoregressive context models are notoriously slow because they often require sequential processing. Furthermore, at very low bitrates, standard mathematical error metrics like Mean Squared Error (MSE) often result in blurry images that lack fine detail. This paper presents a system designed to be both computationally efficient and capable of producing "visually pleasing" results.
Methodology
The system is built on an autoencoder framework with several sophisticated additions designed to enhance both efficiency and perceptual quality.
Nested Hierarchical Hyperprior
Instead of a single hyperprior, the model uses a two-step hierarchy. The latent representation ($y$) is compressed into a hyper-latent ($z$), which is then further compressed into a hyper-hyper-latent ($w$). This allows for extremely efficient entropy coding by better capturing the statistical dependencies in the data.
PixelCNN++Lite and Padé Activation Units
To address the speed issues of autoregressive models, the authors modified the PixelCNN++ architecture. By reducing the number of layers and focusing only on the receptive field of the pixel currently being decoded, they achieved a 10x speedup in decoding time. Additionally, the authors found that Padé Activation Units (PAU) outperformed the traditional Generalised Divisive Normalisation (GDN) in their experiments.
Adversarial and Perceptual Training
To combat the blurriness typical of low-bitrate compression, the model utilizes a Generative Adversarial Network (GAN) objective (WGAN-GP) combined with a VGG perceptual loss. This encourages the reconstruction of realistic textures, even if they are not pixel-perfect matches to the original image.
Results and Discussion
The models were trained and evaluated using the CLIC training dataset and the OpenImage dataset, with validation performed on the Kodak dataset.
Performance and Efficiency
On the CLIC 2020 validation set, the system achieved a PSNR of 29.055 and an MS-SSIM of 0.9502. The architectural optimizations to the context model resulted in a 10x decoding speedup, bringing the model within the strict time limits of the CLIC competition.
Qualitative Improvements
Visual comparisons demonstrate that the combination of VGG and GAN losses effectively removes the "muddy" artifacts common in low-bitrate neural compression. As shown in the comparison below, the model preserves sharp details and realistic textures in complex scenes like forests and urban environments.
Conclusion
This work demonstrates that context-aware learned image compression can be made practical through architectural optimization and perceptual training. The use of nested hyperpriors and PAU activation functions provides a significant performance boost, while the "Lite" context model ensures computational viability. Future work will explore parallelizable context models to further reduce the sequential bottleneck.
Summary updated