Based on The Motion Vector Information for Single-view Depth Map Generation Algorithm

[[abstract]]近年來，人們為了追求立體效果許多3D顯示技術不斷推陳出新。然而，目前市面上的視訊內容還是以2D影像為主流，原因在於3D影像的製作成本昂貴及現存影像依舊以2D影像居多。為達到3D影像之普遍化，除了提昇3D影像製作品質及下降製作成本外，2D影像轉為3D影像系統是目前十分重要的技術。而在轉換技術中，深度圖的品質影響最後的3D效果，因此如何產生出高品質的深度圖將是3D轉換技術中的關鍵。 目前的深度圖產生方法可分為利用單視角影像產生深度圖及雙視角影像產生深度圖本論文進行單視角影像之深度圖產生的方法研究。因為人眼對於動態的物體有較高的興趣，為了在動態物件中提供較佳的3D效果，本論文利用動態向量的資訊來產生深度圖。然而動態向量預估的不準確性將會導致深度圖品質的下降，為了解決此問題，提出以Seed Growing演算法進行物件分割並找尋各物件相對應的特徵點，搭配可調式遮罩的方法增加比對的準確性，另外對前後兩張畫面中相同物件給予相同深度值來解決深度值不連續，用以減少連續畫面中深度值的不連續而造成觀賞者的不舒適性。 在實驗結果的部分，本論文提出之單視角深度圖產生方法在準確率上，與其他方法比較，PSNR的表現上升了0.06~2.93dB，SSIM的評估上升0.01~0.18。另外本論文在主觀性評比上採用了MOS評分，依此觀察，本論文所提出的方法對於畫面當中動量較大的物件有更好的3D效果。[[abstract]]In recent years, researcher develops many 3D technologies for stereo display. However, the content of video is based on 2D images in the market because the manufacturing cost of 3D images is expensive and 2D images exist more than 3D images. To develop 3D technology, the 2D video to 3D video conversion system is the very important technology. Since the quality of the depth map will influences the final 3D effect, the high quality depth map generation is a key scheme in the 2D to 3D conversion system. Now, the depth map generation methods can be classified into Monocular and Binocular depth generation algorithm. In this thesis, the depth map generation method for Monocular video is studied to improve the quality of depth map. The viewers are more interested in the objects which are in motion. In order to get better 3D effect, the information of motion vectors is used to generate the depth map in this thesis. However, the inaccuracy of motion vectors estimation will lead to declining in the quality of the depth map. For solving the problem, this thesis uses Seed Growing for object segmentation and finding corresponding feature points in each object, and it can be combined with an adaptive matching window which is produced by the cross-based matching algorithm to enhance the matching accuracy. To solve the depth value discontinuing and to reduce the reviewer uncomfortable discontinuing depth value in the continuous frames on the proposed algorithm utilizes the information of the object size to track the same object between the current frame and the previous frame, and gives the same depth value to the same object. In the experimental result, it is proved that the result with proposed method is better by comparing the performance, it raises about 0.06 to 2.93dB in PSNR, it raises about 0.01 to 0.18 dB in SSIM. In MOS, the object of the high motion vector can get the better 3D effect in the frame by the proposed algorithm.[[note]]碩