Practical Full Resolution Learned Lossless Image Compression (1811.12817v3)

Abstract: We propose the first practical learned lossless image compression system, L3C, and show that it outperforms the popular engineered codecs, PNG, WebP and JPEG 2000. At the core of our method is a fully parallelizable hierarchical probabilistic model for adaptive entropy coding which is optimized end-to-end for the compression task. In contrast to recent autoregressive discrete probabilistic models such as PixelCNN, our method i) models the image distribution jointly with learned auxiliary representations instead of exclusively modeling the image distribution in RGB space, and ii) only requires three forward-passes to predict all pixel probabilities instead of one for each pixel. As a result, L3C obtains over two orders of magnitude speedups when sampling compared to the fastest PixelCNN variant (Multiscale-PixelCNN). Furthermore, we find that learning the auxiliary representation is crucial and outperforms predefined auxiliary representations such as an RGB pyramid significantly.

An Expert Review of "Practical Full Resolution Learned Lossless Image Compression"

The paper presents a significant advancement in the field of lossless image compression through the introduction of the L3C system, which utilizes deep learning to outperform traditional engineered codecs, including PNG, WebP, and JPEG. This paper describes an innovative approach to entropy coding, using a fully parallelizable hierarchical probabilistic model that is optimized end-to-end specifically for image compression.

Key Contributions

1. Hierarchical Probabilistic Model: At the core of the L3C system is a hierarchical probabilistic model that enables adaptive entropy coding. This model outperforms autoregressive models such as PixelCNN by jointly modeling the image distribution along with learned auxiliary representations. Crucially, it only requires three forward passes to predict all pixel probabilities, offering significant speed advantages.
2. End-to-End Optimization: The proposed model is optimized end-to-end for the compression task. Its parallelizable nature makes it orders of magnitude faster to decode compared to autoregressive models, achieving a $5.31 \times 10^4$ speedup over PixelCNN++ and a $5.06 \times 10^2$ speedup over the multiscale variant, MS-PixelCNN.
3. Learned Auxiliary Representations: The effectiveness of L3C crucially depends on its ability to learn auxiliary representations through a series of feature extractor networks. These learned representations significantly outperform predefined ones, such as an RGB pyramid, in terms of compression efficiency.

Numerical Results

The L3C model’s performance is rigorously evaluated against other lossless compression algorithms. It consistently outperforms PNG, JPEG, and WebP across multiple datasets. While it is only marginally outperformed by FLIF, L3C does not rely on complex hand-engineered techniques, presenting a simpler, learned alternative.

Implications and Future Directions

This paper's contribution lies in demonstrating the feasibility and competitive performance of machine-learned models in the domain of lossless image compression. The L3C model pushes the boundary by showing that learned, parallelizable systems can achieve compression rates and speeds suitable for practical applications. Additionally, this approach lays the groundwork for future developments in domain-specific compression models, which could further leverage the power of learned hierarchical representations.

The theoretical and practical implications are noteworthy. As deep learning continues to evolve, models like L3C could be further refined to utilize even more advanced architectural features or optimization techniques. The prospect of embedding such systems in resource-constrained environments, such as mobile devices, opens new avenues for research and development.

In conclusion, this work marks a pivotal step towards integrating deep learning within traditional image processing and compression techniques, setting a new benchmark for performance and efficiency in lossless image compression. Future exploration may focus on optimizing the system for specific image types or enhancing its performance via hybrid models that combine elements of both learned and traditional techniques.