在高動態場景(high dynamic scene)中,消費型相機無法保存場景的完整亮度範圍。曝光融合(exposure fusion)是一種透過混和不同曝光程度的相片,以完整呈現所有場景細節的技術。近來的曝光融合技術已逐漸走向實際應用面,在此趨勢之下,我們發現大部分的現存方法仍然有一個不切實際的限制:空間域上的對齊限制。為了突破這個限制,我們提出了一個完全不需要空間域上的對齊假設(spatial alignment assumption)或空間域上的對齊前處理(spatial alignment preprocessing)的方法。我們假設我們只能拿到兩張不同曝光的影像,以區域適應場景對比(locally adaptive scene contrast)及曝光度(exposedness)作為混和的判斷標準,將影像在數量平衡直方圖域(number-balanced histogram domain)上混和。為了考慮空間域上的連續性,我們使用馬爾可夫隨機場(Markov random field)來建構問題模型。實驗結果顯示,不論輸入影像有無對齊,我們皆可以得到與現存方法可比擬的結果,而不需要假設輸入影像已對齊或者做任何影像對齊之前處理。
Exposure fusion is a technique for expressing high dynamic scene by fusing differently exposed images. Nowadays, although many recent methods deal with several practical issues, there is still an impractical limitation in the existing exposure fusion methods: the spatial alignment constraint. To overcome this constraint, we propose an exposure fusion method which is totally without spatial alignment assumption or spatial alignment preprocessing. We assume only two input images are available. We use locally adaptive scene contrast and exposedness as fusion criterion to fuse images in the number-balanced histogram domain. To consider spatial continuity, we use Markov Random Field to model our problem. Our experiments demonstrate that our results are comparable with existing methods no matter the input image sequence is aligned or not.