EDAFNet: efficient dual attention fusion network via multi-exposure image for HDR reconstruction
TL;DR
The Efficient Dual Attention Fusion Network (EDAFNet), which combines CBAM-guided initialization, hierarchical local and global attention for cross-exposure alignment and fusion, and a lightweight refinement module based on depthwise separable convolutions, is presented.
Abstract
High Dynamic Range (HDR) reconstruction from multi-exposure Low Dynamic Range (LDR) images requires recovering a wide luminance range while preserving details in bright and dark regions under motion and exposure misalignment. High reconstruction fidelity, however, often comes with increased computational complexity. This paper presents the Efficient Dual Attention Fusion Network (EDAFNet), which combines CBAM-guided initialization, hierarchical local and global attention for cross-exposure alignment and fusion, and a lightweight refinement module based on depthwise separable convolutions. Experiments on the Kalantari dataset, complemented by generalization tests on the Sen, Tursun, and Hu datasets, demonstrate a favorable quality-complexity trade-off. Under the same evaluation protocol across the 15 Kalantari test scenes, EDAFNet achieves the highest linear-domain PSNR and SSIM (43.37 dB and 0.9926, respectively) and the highest mean HDR-VDP-2 score (68.88), while ranking second in the PU domain. A paired non-parametric analysis indicates that the PSNR-l gains are statistically significant against six of the seven evaluated comparison methods after Holm correction. With 0.31M parameters and 0.464T FLOPs, EDAFNet provides strong HDR reconstruction fidelity and competitive perceptual quality while maintaining a compact computational footprint.