影像比對在電腦視覺應用中扮演很重要的角色,而發展一個能準確找到對應點的方法使比對更穩定是關鍵,且比對時有效率也是重點,所以此篇論文提出一個方法能同時提升影像比對的效率與準確度。首先相似的影像用SIFT演算法做分析,經過SIFT技術取出特徵後,我們可以利用此兩組特徵描述子,一組建立最佳k-d樹,另一組用近乎最接近鄰近搜尋法去搜尋最佳候選對應點,此搜尋法可以減少很多比對的時間。最後為了提升準確度,我們利用修改過的區域協調控制矯正方法自動消除錯誤的對應點,同時保留下正確的對應點。我們提出的方法經過測試影像資料庫、歪斜和部分扭曲驗證,和其他方法做比較,實驗結果顯示我們的方法可以同時提升SIFT影像比對的效率與準確度。
Image matching plays an important role in many computer vision applications. It is crucial to develop an accuracy correspondence mechanism and make the matching more stable. The efficiency is also the key course of matching. This thesis presents a model of improving both efficiency and accuracy in image matching. Pictures that have similar contents are first analyzed by SIFT algorithm. After the SIFT feature section, we can utilize two groups of descriptors. A group of descriptors building the optimized k-d tree, and the other group of descriptors finding the best candidate match by approximate nearest neighbor searching and reduce a lot of matching time. Finally, we use a Modifiable Area Harmony Dominating Rectification method to eliminate mismatched key-point couples automatically and protect the matching couples. The proposed method is evaluated on test image database and transformation of shearing effect and thin plate splines. Compared with other method,, the performance of our method is promising in improving the efficiency and accuracy for SIFT image matching.