Abstract
In the past years, learned image compression (LIC) has achieved remarkable performance, with recent methods outperforming VVC in both PSNR and MS-SSIM. However, low bit-rate reconstructions of LIC often suffer from artifacts such as blurring, color drifting, and missing textures, which cause standard metrics to correlate poorly with human perceptual quality. We propose PO-ELIC (Perception-Oriented Efficient Learned Image Coding), adapting the state-of-the-art ELIC model with adversarial training techniques. By applying a mixture of losses—including hinge-form adversarial loss, Charbonnier loss, and a novel patched style loss—we finetune the model for superior perceptual quality. Experimental results demonstrate that PO-ELIC achieves comparable perceptual quality to HiFiC at significantly lower bitrates.
Introduction: The Perception-Distortion Trade-off
Traditional image compression and many modern LIC models typically optimize for Mean Squared Error (MSE) or Peak Signal-to-Noise Ratio (PSNR). While these metrics are mathematically tractable, they often fail to align with human vision. At low bitrates, MSE-optimized models tend to produce "blurry" images because the optimization rewards a safe, smooth average over detailed textures that might be slightly misaligned.
This phenomenon is known as the perception-distortion trade-off: one can achieve high mathematical fidelity (low distortion) or a high-quality visual experience (high perception), but maximizing both simultaneously at low bitrates is extremely challenging. Previous attempts to use Generative Adversarial Networks (GANs) to address this often resulted in unnatural "hallucinated" textures or color drifting. PO-ELIC aims to bridge this gap, providing realistic textures and sharp details without typical GAN artifacts, all while remaining computationally efficient.
Methods: Architecture and Perceptual Optimization
PO-ELIC is built upon the ELIC architecture, known for its "spatial-channel context model" (SCCTX). This allows the model to predict pixel values in parallel, making it significantly faster than older serial models.
To shift ELIC from a distortion-centric to a perception-centric model, several key technical innovations were introduced:
- Adversarial Training (SNGAN): The researchers used a conditional GAN framework with a hinge-form loss and spectral normalization. This setup is more stable than standard GAN training and helps prevent "mode collapse."
- Charbonnier Loss: Instead of standard L1 or L2 loss, a smoother Charbonnier variant was used for reconstruction, providing gentler gradients during training to avoid artifacts.
- Patched Style Loss: A standout feature of the model is the use of style loss calculated locally on 16x16 patches. This forces the model to maintain local texture consistency, accurately reconstructing fine-grained details like butterfly wings or complex textures.
- Multi-Component Objective: The final loss function is a weighted combination of perceptual loss (LPIPS), reconstruction loss, adversarial loss, and style loss.
Key Results
Visual Fidelity and Bitrate Efficiency
PO-ELIC retains clear, realistic details even at bitrates as low as 0.15 bits per pixel (bpp). Qualitative comparisons demonstrate that it preserves significantly more texture in complex structures and dark areas compared to previous state-of-the-art models like HiFiC.
Computational Performance
According to the CLIC 2022 leaderboard, PO-ELIC features the fastest decoder among all learning-based methods, capable of decompressing a 720p image in under 100ms. This proves that parallelized context models are fully compatible with high-quality perceptual reconstruction.
Quantitative Superiority
The model achieves comparable or better scores than HiFiC across a battery of perceptual metrics, including LPIPS, FID (Fréchet Inception Distance), and IQT (Image Quality Transformer), while consistently utilizing a lower bitrate.
Conclusion
PO-ELIC demonstrates a significant leap in the efficiency and quality of perceptual image compression. By combining the high-efficiency ELIC architecture with localized style constraints and stable adversarial training, the authors provide a framework that excels in human-centric visual quality without the computational overhead of serial decoders. While the trade-off with mathematical fidelity metrics like PSNR persists, PO-ELIC represents a practical path forward for high-fidelity, low-bitrate image storage.
Summary updated