立體視覺似人類雙眼系統，藉由對應點比對而得物體深度資訊，但會因遮蔽、影像不連續性、特徵不明顯、光照影響等，造成尋找對應點的錯誤。本論文以可信度傳遞為基礎，利用多次信息傳遞的方式，改進同質區域的視差圖的準確率。由於同質區域內，像素點相似，將導致信息傳遞錯誤，因此，我們提出改良式可信度傳遞於同質區域之立體視覺匹配演算法，第一，我們利用影像梯度結合SAD（sum of absolute differences）資訊判斷以增加可信賴之對應點數目，再使用可信賴對應點傳遞信息，改善信息無法快速從同質區域外部傳遞至內部的問題。第二，我們提出改良式信息傳遞，只需少數的疊代，即能更新出較佳的信息。第三，我們利用加重信息傳遞的權重，加速影像中大範圍同質區域的信息傳遞，達到視差圖快速收斂的程度，以獲得對應準確度較高的視差圖。實驗結果顯示，我們所提出的演算法，在無特徵之平滑區域，時間複雜度較快速可信度傳遞演算法增加了22％，但錯誤率降低了12.2％。 Stereo vision uses two images from different viewpoints to reconstruct the depth of objects. However, stereo correspondence errors may occur due to the object occlusions, depth discontinuities, homogeneous regions and light effects. In this paper, we adopt the belief propagation based algorithm using the message propagation to obtain the disparity map. Due to the similarity of pixels in homogeneous regions, the inaccurate messages may result in the incorrect disparity map. Motivated by this, we further improve the stereo matching algorithm using belief propagation for homogeneous regions. First, we increase the reliable correspondence using the combination of SAD and gradients, and propagate the message from reliable correspondence for homogeneous regions. This improves message propagation from boundary to inside for homogeneous regions. Second, we propose improve message propagation to update optimal message in less iteration. Third, we further accelerate message propagation by increasing the weight of message propagation in larger homogeneous regions. Compared with efficient BP, our experimental results show that our proposed method reduce the inaccurate rate in textureless regions is 12.2％, but increase computation complexity is 22％.
