植基於全域移動向量的自動影像拼接

Abstract

影像拼接是將多張具有重疊視野的相鄰影像組合成一張視野更廣的影像。因為在電腦視覺及電腦圖學上的應用範圍廣泛，像是移動偵測、解析度增強、虛擬實境及機器人導航等，所以一直是很熱門的研究議題。影像間可能存在視差、曝光差及移動物件等問題，這可能會造成鬼影或拼接失敗。本論文提出一個影像拼接法，可以處理這些問題，以避免出現鬼影或拼接失敗。此方法包含影像校準、最佳縫線搜尋及影像混合。在影像校準時，使用尺度不變特徵轉換(SIFT)演算法，找出影像關鍵點並做匹配，然後利用全域移動向量，刪除錯誤的匹配對，最後利用正確的匹配對找出兩影像間的透視變換矩陣。在最佳縫線搜尋時，先利用透視變換矩陣，找出兩影像之重疊區域，然後在重疊區域中搜尋最佳縫線。在影像混合時，先修正影像間的曝光差，再依最佳縫線進行混合。

Image mosaicking is a process of obtaining an image with a wider field-of-view from many images with overlapping fields of view. It has been a hot topic because of the wide range of applications in computer vision and computer graphics, such as motion detection, resolution enhancement, virtual reality, robot navigation. To avoid ghosting effects and mosaicking failure due to moving objects, parallax and exposure differences between images, a method is proposed to solve these problems. The proposed method contains image registration, optimal seam searching, and image blending. In image registration part, SIFT algorithm is used to extract and match SIFT keypoints. Then, the global motion vector is used to eliminate mismatching pairs. Finally, the matching pairs are used to estimate the perspective transformation matrix between two images. In optimal seam searching part, the perspective transformation matrix is used to find the overlapping region of two images and then search the optimal seam in the overlapping region. In image blending part, exposure correction is done for exposure difference between two images. Then, two images are blended based on the optimal seam.