基於亮度與飽和度自適應的多重曝光融合演算法與彈性損失函數設計

Abstract

多重曝光影像融合（Multi-Exposure Fusion, MEF）旨在將不同曝光的低動態範圍（LDR）影像整合為單一視覺平衡的結果，同時保留亮部與暗部的細節。傳統的 MEF 方法多依賴預先定義的規則與多尺度融合策略，然而常因權重估計不穩定與適應性不足，導致光暈（halo）現象與細節保留不完整的問題。 本論文提出一個以優化為核心的 MEF 方法，在可解釋性與適應性之間取得平衡，連結傳統非學習式方法與資料驅動模型。所提出的演算法結合強度–飽和度導向的自適應權重設計與優化導向的金字塔融合架構，並引入曝光區域的自適應分流機制以區分極端與正常曝光區域，提升整體的視覺一致性。在融合階段中，本方法將傳統的金字塔重建轉化為受限優化問題，透過多重損失函數（Composite Loss Function）同時控制光暈抑制、局部對比度與平坦區域的平滑性。 在共計 2,617 組影像序列的實驗中，所提出的方法於 MEF-SSIM 指標上平均表現優於多項代表性傳統與深度學習方法，並在視覺結果上展現更佳的光暈抑制與細節保留平衡。綜合而言，本研究提出一個兼具理論基礎與可調適性的優化式融合框架，為後續混合式或學習型 MEF 模型的發展奠定基礎。

Multi-exposure fusion (MEF) combines differently exposed low dynamic range (LDR) images to produce a single, visually balanced image that preserves both highlight and shadow details. Traditional MEF methods typically rely on predefined heuristics and multi-scale fusion strategies but often suffer from halo artifacts and incomplete detail preservation due to heuristic weighting and limited adaptability. This thesis presents an optimization-based MEF method that bridges traditional non-learning methods with data-driven paradigms through an interpretable and adaptive design. The proposed algorithm integrates intensity–saturation-guided weighting with an optimization-driven pyramid fusion scheme, where an adaptive gating mechanism differentiates extreme and normal exposure regions to enhance perceptual balance. The pyramid fusion stage is reformulated as a constrained optimization problem governed by a composite loss function, enabling explicit control over halo suppression, local contrast, and smoothness. Extensive experiments on a 2,617-sequence benchmark demonstrate that the proposed approach achieves the highest average MEF-SSIM among both traditional and learning-based algorithms. Visual comparisons further confirm its superior trade-off between halo reduction and detail preservation, underscoring the method’s robustness, interpretability, and potential as a foundation for future hybrid or learning-based MEF models.