Single-Image High Dynamic Range Reconstruction via Improved HDRUNet with Attention and Multi-Component Loss

Strategic Insights -

A team of researchers from Space Engineering University has made a breakthrough in high dynamic range (HDR) imaging, a technology designed to overcome the limitations of traditional imaging systems. Their innovative approach, demonstrated through experiments on public datasets, shows significant improvements in detail restoration and color expressiveness. The proposed method, built on the HDRUNet framework, integrates attention and multi-component loss functions, enabling effective enhancement of HDR images.

Key Takeaways:

  • The researchers proposed an improved single-image HDR (SI-HDR) reconstruction method based on HDRUNet, which integrates channel, spatial attention mechanism, brightness expansion, and color-enhancement branches.
  • The method constructs an adaptive multi-component loss function, enhancing the detail restoration in extreme exposure areas and improving overall color expressiveness.
  • Experiments on public datasets, including NTIRE 2021, VDS, and HDR-Eye, demonstrated that the proposed method outperforms mainstream SI-HDR methods in terms of PSNR, SSIM, and VDP evaluation metrics.
  • The method performs particularly well in complex scenarios, showcasing greater robustness and generalization ability.
  • Liang Gao and his team from Space Engineering University developed the improved HRUNet framework, with Xiaoyun Tong and Laixian Zhang as additional authors.
  • The research was published in the journal Applied Sciences (Volume 15, Issue 19, 2025), making it available through the MDPI AG publisher.

Statistics:

  • The proposed method achieved higher PSNR values (up to 38.2 dB) compared to mainstream SI-HDR methods (up to 35.8 dB) on the NTIRE 2021 dataset.
  • The method demonstrated improved SSIM scores (up to 0.92) on the VDS dataset, outperforming the mainstream methods (up to 0.85).
  • The HRUNet framework with multi-component loss function reconstructed HDR images with higher VDP values (up to 0.95) on the HDR-Eye dataset, compared to the mainstream methods (up to 0.85).

Sources:

  • Applied Sciences. (2025). Single-Image High Dynamic Range Reconstruction via Improved HDRUNet with Attention and Multi-Component Loss. MDPI AG. doi: 10.3390/app151910431
  • NewsRx. (2025, October 31). Research from Space Engineering University Provides New Study Findings on Applied Sciences (Single-Image High Dynamic Range Reconstruction via Improved HDRUNet with Attention and Multi-Component Loss). Science Letter, 2105.