- 學位論文
應用雙視角輸入於多視角顯示之虛擬視角產生技術
Multi-views Synthesis Algorithm by Using Two Pairs Color and Depth Images
摘要
近幾年3D影像變得越來越流行，為了增加真實的影像感受，多視角之裸眼顯示器成為新興的3D電視應用，其主要功能是讓不同位置的觀賞者同時受到真實的3D感受且無須配戴眼鏡，但因為拍攝多視角影像有成本過高以及設備架設不易的問題，所以利用DIBR技術進行虛擬影像的產生成為現在主流的虛擬視角合成方式。 一般虛擬影像的創建過程中主要會遇到兩大問題，第一是以相機參數產生不同視角的虛影影像，但相機參數取得不易。另一問題深度圖資訊不連續，或是鄰近像素間視差相異過大，就會使得產生的虛擬影像出現空洞及錯誤輪廓，使其產生的虛擬影像品質下降，進而造成觀賞者的不舒適。 為解決上述遇到的問題，本篇論文提出了應用雙視角輸入於多視角顯示之虛擬視角影像產生技術，利用雙視角的影像資訊創建出符合裸眼電視的多視角虛擬影像，其中應用提出的基於特徵點的立體匹配演算法來計算輸入影像間的真實視差，並藉此視差資訊來取代相機參數，因為降低輸入資訊的限制，所以可以提高多視角合成演算法的便利性。另外為了提升影像品質和觀賞舒適度，最後還加入了錯誤輪廓消除和雙邊濾波器，利用改良後的Inpainting方法針對錯誤輪廓進行消除，再藉由雙邊濾波器降低修補後錯誤輪廓與周圍資訊的色彩差異。最後會將創建出多視角虛擬影像根據裸眼電視的排點方式進行合成輸出。 根據提出之演算法重點將結果分為虛擬視角合成以及瑕疵消除兩塊進行比較，在PSNR部分分別提高了2.67dB和4.57dB，因此根據實驗結果，本論文提出之演算法可以適用於不同裸眼顯示器，並且提升虛擬影像的品質。
並列摘要
In recent years, 3D technologies become more popular. For approaching high perception quality for 3D view experience, multi-view video (MVV) becomes a new 3DTV application. The multi-view TV can bring viewers real 3D experience in the different position simultaneously without 3D glasses. The cost of traditional video capture is too high, to achieve virtual views for multi-view disparity, depth image based rendering (DIBR) algorithm is the popular method. However, there are two problems in the procedure of virtual view generation. One is the projection matrix from camera parameters is usually used to create the virtual view, but the projection matrix is difficult to get. Another problem is artifact effects which include small crack, hole and false contour. Multi-views synthesis algorithm by using two pairs color and depth is proposed in this thesis to solve these problems. The feature based stereo matching is proposed to find the disparity between the input images, and the disparity values are used to replace the projection matrix from camera parameters. To enhance the quality of the generated virtual images and to provide comfortable view experience for viewer, the inpainting based artifact reduction algorithm is proposed to reconstruct the hole and false contour. Then this thesis uses the bilateral filter to smooth the false contour and its surrounding region. Finally, the multi virtual views can be synthesized for multi view display. Experimental results show that the quality of virtual view image in the proposed multi virtual view generation and artifact reduction algorithms can raise 2.67dB and 4.57dB in PSNR, respectively. From the experimental results, the proposed algorithm can provide the high quality multi virtual views for the different multi-view display.