兩階段漸進式多重曝光融合：基於中間曝光生成

Abstract

多重曝光影像融合（MEF）面臨一項根本性的挑戰：現有使用 UNet 或注意力機制的方法必須直接處理亮度差距巨大的極端曝光影像對，這限制了融合品質，特別是在局部區域。本文提出一項新的觀察：當比較的曝光影像之間的相對差異較小時，注意力機制能計算出更可靠的權重。

我們的雙階段框架採用「由全域到局部」的漸進式策略。第一階段（UNet）在網路瓶頸處進行全域特徵融合，將極端曝光的深層特徵合併，以生成亮度均衡的中間曝光影像。此中間影像具有雙重功能：(1) 減少曝光差距以改善注意力權重計算，(2) 作為距離兩個極端曝光等距的參考錨點。第二階段（AMNet）利用該中間影像進行基於注意力的局部細化，在較小的曝光差異下，能更精確地進行特徵比對與選擇性細節恢復。

此架構的分工源於對 MEF 誤差的分析：大多數偽影來自局部曝光失敗，而非全域亮度不平衡。因此，UNet 負責全域協調，而注意力模組則專注於局部增強。我們的解耦式訓練確保每個階段都能針對其特定目標進行最佳化。實驗驗證了我們的理論：中間曝光顯著提升了注意力權重的準確性，在維持全域一致性的同時，於困難區域中達成更優秀的細節保留。

Multi-exposure image fusion (MEF) faces a fundamental challenge: existing methods using UNet or attention mechanisms must directly process extreme exposure pairs with substantial luminance gaps, limiting fusion quality particularly in local regions. We present a novel insight that attention mechanisms compute more reliable weights when comparing exposures with reduced relative differences. Our two-phase framework implements a global-to-local progressive strategy. Phase one (UNet) performs global feature fusion at the network bottleneck, merging deep representations from extreme exposures to generate an intermediate exposure with balanced luminance. This intermediate serves dual purposes: (1) reducing exposure gaps for improved attention computation, and (2) providing a reference anchor equidistant from both extremes. Phase two (AMNet) exploits this intermediate for attention-based local refinement, where smaller exposure differences enable precise feature comparison and selective detail recovery. The architecture division follows MEF error analysis: most artifacts appear as local exposure failures rather than global imbalance. Thus, UNet handles global harmonization while attention targets local enhancement. Our decoupled training ensures each phase optimizes its specific objective. Experiments validate our theory: the intermediate exposure significantly improves attention weight accuracy, leading to superior detail preservation in challenging regions while maintaining global consistency.