克服手振因素之清晰高動態範圍影像合成

Abstract

本論文提出了一個能克服因手振而導致影像模糊的限制，利用手持相機拍攝多張不同曝光時間的照片來合成清晰之高動態範圍影像。 利用多張不同曝光時間的照片來合成清晰之高動態範圍影像是一個十分常見的方法， 然而這些方法卻限制在拍攝這些多張不同曝光時間的照片必須靜止不動，亦即在拍攝時需要腳架或特殊硬體的輔助。 為了能突破這些限制，我們必須克服手持相機所產生的兩個主要問題：第一，照片間的對齊問題，第二，拍攝的照片容易因手振導致模糊，尤其在曝光時間愈長時，情況愈嚴重。 為克服這些問題，我們針對不同曝光時間的照片先做對齊後，並提出一個新的演算法運用這多張可能有模糊的影像來合成清晰之高動態範圍影像。 在我們提出的演算法中，我們假設這多張不同曝光時間的照片都是由同一張清晰的高動態範圍影像但經過不同的 晃動路徑及曝光時間而成。利用這多張影像的資訊，運用貝氏定理定義出問題的機率模型，透過反覆地解出每張影像對應的晃動路徑， 並反覆地更新合成出來的高動態範圍影像，最後更新相機的函數，依此不斷地重複更新此三項參數直到演算法收斂為止。 最後我們將可求得令人滿意且擁有豐富及清楚細節的高動態範圍影像。 為證實我們演算法的有效性，我們將對合成的影像以及實際拍攝的影像進行測試，並跟其他兩個相關的方法做比較。

In this thesis, we present a novel technique to reconstruct a high quality high dynamic range image from a set of differently exposed and possibly blurred images from a hand-held camera. Compositing high dynamic range image from differently exposed photographs is a common approach, however, it often requires a tripod to keep the camera still when taking photographs of different exposures. To ease the process, one often prefer to use hand-held cameras. This, however, leads to two problems, misaligned photographs and blurred long-exposed photographs. To overcome these problems, we adapt an alignment method and propose a method for HDR reconstruction from possibly blurred images. We conqure the blurry artifact caused by camera shake from hand-held cameras to reconstruct a clear high dynamic range image. Instead of applying the naive solution that separately deblur each photograph and then composite these deblurred results to construct a high dynamic range image, we seek to utilize the correlation among blurred images, based on the assumption that all observations come from the same latent high dynamic range image convolved by different blur kernels. We adopt a Bayesian approach to define our probability model and apply maximum-likelihood approach to iteratively perform kernel estimation, high dynamic range image reconstruction and camera curve recovery. Until convergence, we obtain an satisfying high dynamic range image with rich and clear structure. To demonstrate the effectiveness of our algorithm, we tested our implementation on both synthetic examples and real photographs and compared favorably to two related methods.