- 學位論文
基於深度圖之影像處理及其應用
Depth-Based Image Processing and Its Applications
摘要
近幾年，隨著立體匹配演算法的日趨成熟與深度相機(如: Kinect)的快速發展下，與深度圖有關之研究及其應用在影像處理領域裡漸漸受到高度的關注。為了讓基於深度圖之影像處理領域能更完整與豐富，在此博士論文中，我們提出多項先進的基於深度圖之影像處理技術與應用。其中包括利用深度合成無錯模組來實現3D 不可視顯性浮水印技術，深度輔助邊緣檢測演算法，深度圖空洞修補以及供視障用戶使用的聽覺深度圖像系統。深度圖是立體影像生成的主要輸入訊號，所以深度訊息對於立體影像合成的品質有很大的影響。本論文提出的3D 不可視顯性浮水印技術是利用深度合成無錯模組將欲傳遞之輔助信息藏入深度圖內並與3D 影像內容結合，進而不影響合成後的立體影像品質。此外，我們也將此方法也擴展到多視角合成的不可視顯性浮水印。實驗結果顯示，該方法具有較強的穩健性、隱蔽性且完全不影響合成後的立體影像品質。 深度相機因為其低廉的價格而被廣泛使用。但是目前從深度相機得到原始的深度圖都有破洞、雜訊、等問題。其中邊緣不完整問題尤其嚴重。因此，利用深度圖前應先將其做破洞修補。本論文提出了一套利用深度輔助邊緣檢測演算法來進行深度圖像修補技術。只有在邊緣同一側的像素被選為參考像素，再通過平面擬合方法來修補破洞。此方法，能使邊緣附近的模糊效果最小化，因為在邊緣異側的相鄰像素不會被列入參考像素。實驗結果顯示，此方法可以產生無破洞、更精確且邊緣更完整的深度圖。 最後，我們還提出了一套供視障用戶以不同的方式實現可視化圖像的聽覺深度圖像系統。
並列摘要
Many depth–based applications have gained increasing interest in recent years due to the advances in stereo matching algorithms and depth camera sensor (e.g., Kinect). The use of depth maps facilitates many difficult computer vision tasks and assists the generation of stereoscopic views in Three Dimensional TeleVision (3DTV). In order to achieve a better and complete depth–based image processing field, a number of advanced depth–based applications and image processing technologies are proposed in this dissertation, including the 3D unseen visible watermarking (UVW) using depth no synthesis error (D-NOSE) model, depth-assisted edge detection algorithm, depth hole filling and auditory depth images system for visually impaired users. Depth information is a main input in view synthesis and the quality of synthesized views is very sensitive to the depth information. Therefore, we proposed a 3D UVW scheme based on D-NOSE model for putting the auxiliary information into depth map and jointing the 3D multimedia content together without affecting the quality of synthesized views. Furthermore, this 3D UVW scheme is also extended to multi-view UVW. Experimental results show that our proposed method have strong robustness 、imperceptibility and the quality of synthesized views is totally unaffected. Depth camera is now in widespread use since its low price. However, the quality of the depth map captured by depth camera is degraded by various defects, especially near object boundaries. Hence, before using Kinect depth data, these invalid regions should be filled first. We proposed an efficient hole filling strategy based on a new depth-assisted edge detection algorithm. Only pixels which are on the same side of the edges are selected as reference pixels to predict missing pixels by using plane fitting method. In this way, neighbor pixels on the opposite side are not considered as reference pixels so that the blurring effect near edges is minified. Experimental results demonstrate that the proposed hole filling strategy can generate accurate depth map and has better performance of depth map accuracy than other existing methods. Furthermore, an auditory depth images system is also introduced which can provide a different way to visualize the image for visually impaired users.