TL;DR
This paper introduces a novel deep learning approach for lossless image compression by adapting the "hyperprior" concept from lossy compression and employing flexible Gaussian Mixture Likelihoods for modeling image data. The proposed method achieves state-of-the-art performance among learned lossless compression techniques and outperforms established codecs like JPEG2000 and WebP on JPG images.
What Problem Does This Paper Address?
Traditional image compression methods, while effective, often rely on hand-crafted components and may not be optimal for the ever-increasing demands of high-resolution imagery and diverse multimedia applications. While deep learning has shown great promise in lossy image compression, its application to lossless compression has been less explored. Lossless compression is crucial for applications where perfect reconstruction is paramount, such as medical imaging or archival purposes. The challenge lies in accurately modeling the complex probability distributions of image data to achieve high compression ratios without any loss of information. Existing learned lossless methods, like L3C, have made progress, but there's still room for improvement in terms of compression efficiency and training stability.
Methods
The core of this work lies in two main innovations:
- Generalizing the Hyperprior for Lossless Compression: The "hyperprior" concept, previously used in lossy compression to model spatial dependencies in latent representations, is adapted for lossless compression. This involves introducing a secondary, compressed representation (the hyperprior) that helps model the main latent representation.
- L2-Norm for Faster Training: To accelerate the training of these complex neural networks, the authors introduce an L2-norm term into the loss function. This term directly penalizes the difference between the ground-truth pixel values and the predicted mean values of the latent representations, providing a more stable and faster convergence during training.
- Discretized Gaussian Mixture Likelihoods: Instead of relying on simpler distributions like univariate Gaussian or logistic, the paper proposes using Gaussian Mixture Models (GMMs) to represent the probability distributions of the latent codes. This offers greater flexibility in modeling the complex, non-Gaussian marginal distributions observed in image data, leading to more accurate probability estimations and thus better compression. The GMM parameters (weights, means, and variances) are learned by the network.
Datasets
- Training: The model is trained on approximately 35,000 patches (128x128 pixels) from the ImageNet ILSVRC dataset.
- Evaluation: Performance is evaluated on three datasets:
- CLIC professional validation dataset (41 images)
- CLIC mobile validation dataset (61 images)
- Kodak dataset (24 images)
- For evaluation, the test images are converted to JPG with quality level 95 (following L3C's approach) and also evaluated as lossless PNGs to highlight the impact of training data artifacts.
Key Insights and Ideas
- Hyperprior's Versatility: The hyperprior, initially designed for lossy compression, proves to be a valuable tool for lossless compression by effectively capturing spatial redundancies.
- L2-Norm as a Training Stabilizer: The L2-norm term is a simple yet effective way to speed up training and improve stability in learned lossless compression models, addressing a common challenge in this field.
- GMMs for Better Modeling: The observation that image data distributions are not strictly Gaussian is critical. Using GMMs allows the model to adapt to these complex distributions, leading to improved compression performance.
- JPG Artifacts Impact: The paper highlights how artifacts in JPG training data can affect the performance of learned models when evaluated on lossless formats, suggesting a need for more robust training strategies or lossless training datasets.
Key Results
- The proposed method outperforms existing learned lossless compression approaches, specifically L3C, on the CLICP and Kodak datasets.
- It also achieves better compression performance than JPEG2000 and WebP on JPG images across all tested datasets.
- While slightly worse than L3C on the CLICM dataset for JPG images, the proposed method generally shows competitive or superior results.
- The paper demonstrates that Gaussian distribution, among univariate models, performs best for modeling the latent codes, but GMMs offer further improvements.
Limitations and Future Work
- JPG Artifacts: The authors acknowledge that the performance gap on lossless PNG images compared to JPG images is due to artifacts in the JPG training data.
- Future Work:
- Investigating methods to mitigate the impact of compression artifacts in training images.
- Exploring the creation of lossless image datasets for training.
- Further research into optimal distribution models for learned lossless compression.
Summary updated