Boosting HDR Image Reconstruction via Semantic Knowledge Transfer.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Recovering High Dynamic Range (HDR) images from Standard Dynamic Range (SDR) images is difficult, especially with degraded or incomplete SDR data.
• Scene-specific semantic priors can aid restoration but face domain/format gaps when applied to HDR imaging.

Purpose of the Study:

• To propose a general framework for transferring SDR semantic knowledge to boost HDR reconstruction.
• To address the domain/format gap challenge in applying SDR semantic priors to HDR imaging.

Main Methods:

• Introduced the Semantic Priors Guided Reconstruction Model (SPGRM) to leverage SDR semantic knowledge for HDR reconstruction.
• Employed a self-distillation mechanism to align color and content information using semantic knowledge.
• Utilized a Semantic Knowledge Alignment Module (SKAM) to transfer internal feature semantic knowledge and fill missing content.

Main Results:

• The proposed framework significantly enhances HDR imaging quality.
• The method effectively boosts existing HDR reconstruction techniques without altering their network architecture.
• Demonstrated successful transfer of semantic knowledge across SDR and HDR domains.

Conclusions:

• The developed framework offers a robust solution for improving HDR reconstruction from degraded SDR images.
• Semantic knowledge transfer via self-distillation and SKAM is effective in overcoming domain gaps.
• The approach provides a generalizable method to enhance various HDR reconstruction models.