影像處理在今天已經是多媒體領域的一門重要學問,應用在立體影像上的立體匹演算法亦扮演著重要的地位。本論文主要是先介紹與立體匹配演算法有關的數學模型以及相關假設,再來利用影像分割的原理,討論影像中深度平面與原始影像信號中相關性的區域,分析關聯性決定出區域,最後再合併區域找尋最佳解。在此使用區域正交相關匹配深度演算法(Local matching with Orthogonal Integral Segment)配合簡單樹狀動態演算法(Simple tree dynamic programming )來決定初始深度影像,並利用正向匹配與反向匹配找出可信點區域後,細部修正可信點區域使之雜訊大幅減少,並對原始影像進行顏色相關度切割決定出深度平面,再利用平面方程式與B-spline平面修正使預測的深度平面可以有不同曲度展現形式。接下來利用原始影像與經過深度修正後的影像進行色彩統計上的相似度比較,並佐以比對Census轉換後的紋理影像,找出最適合的影像分割與曲度集合,最後對深度影像不連續區域做以正交相關區域為基底,進行匹配能量重定修正藉以達到更好的表現。整體演算法於Middle Bury立體視覺影像研究資料庫所提供之四組影像對計算結果與真實深度之比較,錯誤率分別為Tsukuba 2.37%、Venus 1.10%、Teddy 7.83%、Cones 3.99% 整體錯誤排名為56名
Nowadays, 3D image processing has become a critical technique in multimedia applications. Particularly, stereo matching plays an essential role where a segmentation-based stereo matching approach is widely adopted in disparity computing, region clustering, and segmentation refining. This work first introduces the theoretical foundation with respect to segmentation-based stereo matching. Second, based on image segmentation, we explore the depths, disparity planes and relevance among segmented regions and then determine the association to find an optimized solution of region grouping. Particularly, the local matching with orthogonal integral segmentation and simple tree dynamic programming are used to determine initial depths of planes. Additionally, the B-spline surface, plane fitting function and majority voting are applied to predict reliable region depths where the cost function is based on statistical color similarity using the mutual information and texture similarity using the census transform. Third, the edge correction scheme based on the orthogonal integral segmentation is adopted to raise the performance. Finally, the obtained depths are used as an initial depth map for the next iteration. When the cost function cannot be lowered, the iteration is stopped. The four sample image sets from Middlebury are tested by using the proposed method to yield the error rates of bad pixels, 2.37%, 1.10%, 7.83% and 3.99% for Tsukuba, Venus, Teddy and Cones, respectively, where the overall ranking score is 56.