LDP-MEF: Lossless Detail Preservation Multi-Exposure Image Fusion Network via Multi-Attention Cooperative Guidance

Abstract

Multiexposure image fusion (MEF) is aimed at generating a well-exposed fused image from low-dynamic-range images captured at different exposures, thereby supporting HDR-oriented imaging goals. Existing deep MEF networks often employ multiscale architectures to enlarge receptive fields; however, conventional downsampling and upsampling inevitably cause irreversible detail loss. In addition, deeper feature extraction tends to accumulate redundant or detrimental features, whereas simple fusion of global and local features weakens local illumination correction. To address these issues, we propose a lossless detail-preserving network for MEF, termed LDP-MEF. First, a Haar wavelet feature extraction module (HFEM) employs reversible Haar encoding and decoding to retain all four subbands during resolution conversion and enhances high-frequency subbands to preserve fine textures in extreme exposure regions. Second, a multiattention-guided (MAG) fusion module jointly performs self-channel gating, cross-channel allocation, and cross-spatial allocation to suppress detrimental features and achieve improved color-consistent fusion. Third, an HDR image generation module (HIGM) injects global exposure cues into local patch generation to enable adaptive local illumination correction. Experiments on the SICE dataset show that LDP-MEF outperforms sixteen state-of-the-art methods. Relative to the best competing results, it reduces the LPIPS score by 2.2% and improves the CC by 2.0% while providing better texture preservation, more faithful color reproduction, and the best values across all eight metrics.