AFUNet: Cross-Iterative Alignment-Fusion Synergy for HDR Reconstruction via Deep Unfolding Paradigm

2025-06-30

AI Summary

Existing learning-based HDR reconstruction methods heavily rely on empirical architectural design and lack theoretical foundations. To address this, we propose a deep-unfolding-based co-optimization framework that jointly performs alignment and fusion. Starting from maximum a posteriori (MAP) estimation, we explicitly model and unfold the conventional iterative optimization process into a learnable neural network, wherein each layer alternately executes spatial alignment and channel-wise fusion to achieve progressive joint optimization. This framework ensures theoretical rigor and structural interpretability, eliminating black-box design. Evaluated on multiple standard HDR benchmarks, our method surpasses state-of-the-art approaches in PSNR, SSIM, and perceptual quality, significantly improving highlight/shadow detail recovery and dynamic range consistency.

Technology Category

• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search
• Machine Learning: Learning with Manifolds

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Existing learning-based methods effectively reconstruct HDR images from multi-exposure LDR inputs with extended dynamic range and improved detail, but they rely more on empirical design rather than theoretical foundation, which can impact their reliability. To address these limitations, we propose the cross-iterative Alignment and Fusion deep Unfolding Network (AFUNet), where HDR reconstruction is systematically decoupled into two interleaved subtasks -- alignment and fusion -- optimized through alternating refinement, achieving synergy between the two subtasks to enhance the overall performance. Our method formulates multi-exposure HDR reconstruction from a Maximum A Posteriori (MAP) estimation perspective, explicitly incorporating spatial correspondence priors across LDR images and naturally bridging the alignment and fusion subproblems through joint constraints. Building on the mathematical foundation, we reimagine traditional iterative optimization through unfolding -- transforming the conventional solution process into an end-to-end trainable AFUNet with carefully designed modules that work progressively. Specifically, each iteration of AFUNet incorporates an Alignment-Fusion Module (AFM) that alternates between a Spatial Alignment Module (SAM) for alignment and a Channel Fusion Module (CFM) for adaptive feature fusion, progressively bridging misaligned content and exposure discrepancies. Extensive qualitative and quantitative evaluations demonstrate AFUNet's superior performance, consistently surpassing state-of-the-art methods. Our code is available at: https://github.com/eezkni/AFUNet

Problem

• Reconstruct HDR images from multi-exposure LDR inputs
• Decouple HDR reconstruction into alignment and fusion subtasks
• Bridge misaligned content and exposure discrepancies progressively

Innovation

• Deep unfolding for HDR reconstruction synergy
• Alignment-Fusion Module with alternating refinement
• MAP estimation with spatial correspondence priors