具有疊代定位及變動取樣頻率的環場影像鑲嵌

Abstract

在現代化的資訊世界，數位多媒體影音與網路緊密的融合搭配是必然的趨勢。環場影像鑲嵌(panoramic image mosaicing)則是在各種數位影像相關應用中，最基礎的技術之一。本篇論文提出一套以方便使用者操作為前提，避免需要使用者介入調整的環場影像鑲嵌系統，希望將一連串以手持不定焦距的相機朝任意角度旋轉拍攝的影像，在盡量減小相機位移的條件下，自動化的建立出可供連續性轉動瀏覽的環場影像。 我們由投影轉換系統(projective transformation system)為基礎，經過相機校正與影像轉換矩陣的求取，在考慮相鄰影像曝光值可能改變的情況後，以亮度調整(blending)的運算求得各投影射線亮度值，最後建立整個環場鑲嵌影像。由於輸入的影像是由手持相機拍攝，故可能有些許位移而會使鑲嵌影像產生裂口(gap)現象。我們提出一套三階段以Levenberg – Marquardt 演算法為基礎的修正程序，以消除影像疊合後會出現的誤差。 最後，透過偵測影像是否變焦拍攝，提出自動化調整投影射線取樣頻率的技術，以期能以最小的儲存空間，記錄下最佳的影像品質。

In the information age, the interplay of multimedia and Internet is becoming more and more important. Panoramic image mosaicing is a fundamental technique in digital image-based rendering applications. This thesis presents an approach to facilitating the seamless stitching of a sequence of overlapping images taken with a panning camera, hold by hands, whose lens center remains relative still, but focal length varying. We apply a three-phase projective transformation process to register the sequence of images in a common coordinate system. And we perform the intensity blending to remove the possible uneven illumination conditions in the different images. The above three-phase registration process iteratively eliminates the ghost or gap phenomenon in the final mosaicing output. Finally, we can check the value of the camera focal length to decide the sampling rate for warping the image projection rays to a cylinder or a sphere. In this way, we can produce a good image mosaicing result with a reasonable size of memory storage.