TL;DR
This paper introduces BPGAN, a novel framework for compressing both images and speech using Generative Adversarial Networks (GANs). By iteratively searching for optimal latent representations through back-propagation and employing ADMM for efficient quantization, BPGAN achieves superior compression quality across various metrics compared to existing methods.
What Problem Does This Paper Address?
Efficiently compressing data is crucial for managing the ever-increasing volume of information in our digital world, especially for bandwidth-constrained applications like the Internet of Things. Traditional compression methods often struggle to balance compression ratio with signal quality, particularly for complex data like images and speech. While deep learning approaches have shown promise, they often require separate models for different signal types and can fail to preserve fine details at very low bitrates. This paper aims to bridge this gap by proposing a unified, GAN-based compression framework that can handle diverse signal types while achieving high fidelity at significantly reduced data rates.
Methods
The core of BPGAN lies in its innovative compression methodology. Instead of relying on a traditional encoder-decoder architecture, BPGAN uses a pre-trained GAN generator. The compression process involves iteratively searching for an optimal latent vector (z) that, when fed into the generator (G), produces a signal (G(z)) that closely matches the original target signal (x) according to a defined loss function (F(x, G(z))). This search is performed via back-propagation, updating only the latent vector, not the generator's weights.
To further enhance compression, the framework incorporates Alternating Direction Method of Multipliers (ADMM) for non-uniform quantization of the latent vectors. This is crucial for reducing the bit rate. The optimization problem is formulated to minimize the loss function while ensuring the latent vector is quantized. ADMM is used to solve this non-convex optimization problem by iteratively updating the latent vector, auxiliary variables, and Lagrange multipliers. Finally, lossless Huffman coding is applied to the quantized latent vectors for further compression.
The training methodology involves two stages: first, training a standard GAN for the target signal type, and second, retraining the GAN with quantized latent vectors to improve its performance in the quantized space.
Datasets
- Image Compression:
- Training: Open Images Dataset V5 (9 million images)
- Evaluation (PSNR/MS-SSIM): Kodak dataset
- Evaluation (Classification): ImageNet dataset
- Speech Compression:
- Training/Evaluation: TIMIT dataset (6300 sentences from 630 speakers)
The paper uses standard datasets for image and speech, but the novelty lies in how these datasets are utilized within the BPGAN framework for evaluating compression performance.
Key Insights and Ideas
- Unified Framework: The ability to use a single GAN-based framework for both image and speech compression is a significant step towards more generalized data compression.
- Back-propagation for Latent Search: Using iterative back-propagation to "search" for the optimal latent representation is a clever way to leverage the generative power of GANs for compression, allowing for objective-aware compression tailored to specific tasks (e.g., signal classification).
- ADMM for Efficient Quantization: The integration of ADMM for non-uniform quantization is key to achieving high compression ratios without sacrificing too much quality. It effectively handles the non-convexity introduced by quantization.
- Objective-Aware Compression: The framework's flexibility in using different loss functions allows for compression optimized for specific downstream tasks, such as improving classification accuracy on decompressed images.
Key Results
- Image Compression: BPGAN achieved superior PSNR and MS-SSIM scores compared to GAN-based [5], JPEG, and BPG at similar bitrates (around 0.286-0.306 bpp). Crucially, BPGAN also demonstrated significantly better ImageNet classification accuracy on decompressed images, indicating better preservation of actual details.
- Speech Compression: At a very low bitrate of 2 kbps, BPGAN achieved a better MUSHRA subjective quality score than other methods, even those with higher bitrates (except Opus at 9 kbps). While PESQ scores were similar across methods, BPGAN showed a significantly lower phoneme error rate (PER) in recognition tests.
- Parameter Sensitivity: Figure 4 illustrates the tradeoff between bit rate and quality, showing how adjusting latent vector dimensions and quantization levels allows for fine-tuning the compression performance.
Limitations and Future Work
- Computational Cost: While the paper mentions accelerating the back-propagation process by initializing latent vectors with an encoder output, iterative back-propagation can still be computationally intensive.
- GAN Training Complexity: Training GANs, especially for high-resolution signals, can be challenging and require significant computational resources.
- Subjective Quality: While objective metrics and classification accuracy are strong indicators, subjective perceptual quality can still be a challenge to fully capture. The paper notes that GAN-based methods can sometimes synthesize "realistic" but inaccurate details.
Future work could explore more efficient back-propagation techniques, investigate GAN architectures better suited for extremely low bitrates, and further refine the loss functions to better align with human perception of quality.
Summary updated