HPGN: Hybrid Priors-Guided Network for Compressed Low-Light Image Enhancement

Technical Summary: HPGN for Compressed Low-Light Image Enhancement

Problem Statement
In practical applications, low-light images are frequently compressed to reduce storage and transmission costs. However, this compression introduces artifacts (e.g., blocking, blurring) that compound with the inherent degradation of low-light conditions. Existing Low-Light Image Enhancement (LLIE) methods typically fail to mitigate these compression artifacts when applied directly to compressed inputs. Furthermore, current approaches to Compressed Low-Light Image Enhancement (CLLIE) often rely on a two-stage strategy: first deblocking the image, then enhancing illumination. This cascaded approach leads to error accumulation, increased model complexity, and significant training time costs, as separate models are often required for different JPEG Quality Factors (QF). Additionally, many existing methods do not explicitly utilize compression metadata (such as QF or Quantization Matrices) to guide the enhancement process.

Methodology
The authors propose the Hybrid Priors-Guided Network (HPGN), a one-stage framework designed to jointly enhance compressed low-light images across varying compression levels. The architecture consists of two primary components:

1. Hybrid Information Filter (HIF): This is a plug-and-play module that dynamically adjusts features by integrating three types of priors:

   • Illumination Prior: Derived from the input compressed image ($I_{comp}$) via a global brightness estimation and local illumination feature extraction process.
   • Compression Priors:
     • Quality Factor (QF) Branch: Utilizes the QF to generate mapping coefficients ($M_{qf1}, M_{qf2}$) that perform channel-wise multiplication on illumination features. This addresses channel correlation changes and feature shifts caused by quantization.
     • Quantization Matrix (QM) Branch: Utilizes the DCT Quantization Matrix to generate spatial mapping coefficients ($N_{qm}$). These are applied via element-wise multiplication to model spatial feature distributions, amplifying enhancement in key areas (like textures and edges) and locally correcting regions with significant quantization degradation.
   • The HIF combines these priors to produce refined features that guide the subsequent enhancement.

2. Image Enhancer (IE): A CNN-based module composed of stacked Recursive Multi-scale Residual Blocks (RMRB). Each RMRB contains Multi-scale Residual Blocks (MRB) that employ a multi-scale branch architecture. Input features are downsampled to capture correlations at different resolutions, processed through Context Blocks (CB), and dynamically fused. This structure aims to handle varying light distribution scales while maintaining low complexity.

Training Strategy
To enable a single model to handle images with varying compression qualities, the authors employ a random QF generation strategy during training. Instead of training separate models for specific QFs, the network is trained on images compressed with randomly generated QFs. This allows the model to learn a unified representation for diverse compression levels, significantly reducing training time and computational resources compared to multi-model or two-stage approaches.

Key Contributions

• One-Stage Framework: HPGN is proposed as a unified, one-stage network for compressed low-light enhancement, eliminating the need for cascaded deblocking and enhancement models.
• Hybrid Information Filter (HIF): A novel module that leverages compression priors (QF and QM) alongside illumination priors to mitigate information loss during brightness adjustment. The authors highlight its "plug-and-play" nature, allowing it to be embedded into other LLIE methods.
• Random QF Training: A strategy that enables a single model to robustly enhance images across varying compression levels, addressing the scalability issues of previous methods that require specific optimization for each QF.

Experimental Results
The method was evaluated on four datasets: LOLv1, LOLv2-real, LOLv2-syn, and a custom LOLv1-randomQF dataset.

• Quantitative Performance: On the LOLv1-randomQF dataset, HPGN achieved state-of-the-art performance, outperforming the second-best approach (CAPformer) by up to 0.21 dB in PSNR and achieving higher SSIM scores.
• Efficiency: HPGN demonstrates superior efficiency, requiring only 2.69M parameters, which is approximately 30.7% of the parameters used by CAPformer (8.77M).
• Scalability: Ablation studies and integration tests (Table 2) show that adding the HIF module to existing LLIE frameworks (MIRNetv2 and Retinexformer) significantly improves their performance on joint tasks with minimal parameter increase.
• Visual Quality: Qualitative comparisons indicate that HPGN produces superior visual results with reduced noise and fewer compression artifacts compared to both standard LLIE methods and two-stage cascaded approaches.

Significance and Claims
The paper claims that HPGN addresses the critical gap in handling compressed low-light images by unifying compression artifact removal and illumination enhancement into a single, efficient stage. By explicitly utilizing compression metadata (QF and QM) as guidance, the method adapts to varying compression levels without retraining. The authors position HPGN as a versatile solution that not only outperforms existing SOTA methods in accuracy but also offers a more computationally efficient and flexible framework for real-world applications where image compression levels vary. The paper notes that while the current focus is on JPEG, the core concept of using compression metadata to guide enhancement is extensible to other standards, though specific adaptations would be required for non-JPEG or extreme compression scenarios.