摘要
在過去高動態範圍影像是經由一系列不同曝光時間的影像來合成，可用合成方式的不同分為兩類，第一類是由[1]所提出的高動態，轉換至輻射照度域(radiance domain)來合成，第二類是由[2]所提出的曝光融合，在像素域(pixel domain)經由對比度、飽和度、良好曝光度來權重一系列不同曝光時間影像，以合成高動態範圍影像。在過去的方法大多數是假設輸入影像已經完美對齊，然而在實際拍攝場景時，輸入影像並沒有辦法完全對齊。在本篇論文系統中經由dual-camera方法來分別得到自動曝光(automatic exposure)與補償曝光(complimentary exposure)兩張不同曝光與視角的影像，在過去影像對齊技術使用尺度不變特徵轉換(Scale-invariant feature transform)，可以在相同曝光時間影像中找到特徵點並轉換，但是此方法會因為不同曝光時間影像，難以找到對應特徵點而導致不對齊結果，因此本篇論文提出新的對應方法，利用區域亮度分佈一致性克服特徵點難以對應問題，又因先前方法利用全域對應特徵點搜尋，導致運算時間增加而降低演算法效率，在此本篇論文提出相機幾何對應關係搭配提出的對應方法，縮小對應搜尋範圍，有效地提升演算法效率，解決對齊問題以達到高動態範圍影像合成。並且在實驗結果章節，展示比較本篇論文與之前不同對齊方法結果與運算時間。
並列摘要
Most of existing methods acquired high quality images from a sequence of differently exposed images that can be differentiated either high dynamic range (HDR) [1] or exposure fusion [2]. Most of them are focused on global HDR image generation for static scenes and assume the input images are perfectly aligned. However, these methods are not effective to cope with practical scenario. This paper presents an effective and real-time ROI-base image alignment technique for two different exposed images automatic-exposed and complementary-exposed with different viewpoint. In previous work, SIFT (Scale-invariant feature transform) can accurately match the feature points from two different images with the same exposure time. But it will be hard to match the correspondence feature points in HDR synthesis system due to the different exposed images. In this paper, we propose a matching method that overcomes the cause of exposure limits by taking advantage of consistent luminance distributions. In previous work, it is well-known with great time consumption because of global alignment. Therefore we utilize the camera geometric relations with our matching method to replace the global alignment in order to decrease the candidates of matching region. Our experiments demonstrate that our results are effective and high-quality image alignment for HDR synthesis system.