Boosting HDR Image Reconstruction via Semantic Knowledge Transfer

TL;DR

This work tackles the ill-posed problem of reconstructing HDR images from degraded SDR exposures by introducing a semantic-prior based framework that transfers knowledge from the SDR domain to HDR without modifying existing HDR models. The approach hinges on a Semantic Priors Guided Reconstruction Model (SPGRM) that leverages a Semantic Knowledge Bank (SKB) derived from fused HDR and a Semantic Knowledge Transfer Scheme (SKTS) to distill semantic information back into the Original Reconstruction Model (ORM). A Semantic Knowledge Alignment Module (SKAM) facilitates feature-level alignment in a shared latent space, enabling progressive, masked knowledge transfer. Experiments across RAW and sRGB HDR datasets show consistent improvements over multiple baselines, including state-of-the-art methods, with robust qualitative gains in texture, color stability, and ghosting reduction. The method achieves these gains without changing network architectures, albeit with increased training cost due to SPGRM and semantic preprocessing, and points to future work on broader color spaces and lighter knowledge-transfer strategies.

Abstract

Recovering High Dynamic Range (HDR) images from multiple Standard Dynamic Range (SDR) images become challenging when the SDR images exhibit noticeable degradation and missing content. Leveraging scene-specific semantic priors offers a promising solution for restoring heavily degraded regions. However, these priors are typically extracted from sRGB SDR images, the domain/format gap poses a significant challenge when applying it to HDR imaging. To address this issue, we propose a general framework that transfers semantic knowledge derived from SDR domain via self-distillation to boost existing HDR reconstruction. Specifically, the proposed framework first introduces the Semantic Priors Guided Reconstruction Model (SPGRM), which leverages SDR image semantic knowledge to address ill-posed problems in the initial HDR reconstruction results. Subsequently, we leverage a self-distillation mechanism that constrains the color and content information with semantic knowledge, aligning the external outputs between the baseline and SPGRM. Furthermore, to transfer the semantic knowledge of the internal features, we utilize a Semantic Knowledge Alignment Module (SKAM) to fill the missing semantic contents with the complementary masks. Extensive experiments demonstrate that our framework significantly boosts HDR imaging quality for existing methods without altering the network architecture.

Figures (8)

• Figure 1: Performance comparison on the BracketIRE dataset, which consists of RAW format data pairs, demonstrates that our framework enables the baseline model to achieve significant improvements in PSNR (pixel accuracy), SSIM (detail preservation), and LPIPS (perceptual fidelity) without altering the network architecture.
• Figure 2: The Proposed Semantic Knowledge Transfer Framework. During training, the SPGRM enhances ORM's initial HDR results by integrating semantic priors. Meanwhile, the Semantic Knowledge Transfer scheme boosts the performance of the ORM through model/feature-level self-distillation. During inference, only the optimized ORM is deployed, ensuring high efficiency.
• Figure 3: The process of Semantic Knowledge Alignment Module, which conducts semantic knowledge transfer at the feature level.
• Figure 4: The process of the training and inference stages.
• Figure 5: Visual comparison of baseline methods with and without our framework on Tel and Kalantari’s datasets.