目前市面上的三維內容產生器並無法直接取像產生立體影像,都必須先截取實景中的2D影像後,依靠軟體做連結重新建構成立體影像,而一張立體影像的品質好壞,軟體於其中佔有一個重要的地位。 方筒鏡成像系統是一個特殊設計的三維內容產生器,能夠於實景中取像後產生一張同時具有九個視場的二維影像,稱為方筒鏡影像,而3D重建時所需要的各項重要資訊,例如:深度,則隱含於這九個視場中,有效地將各個視場獨立分割出來便可以實行立體影像的各項應用。 本論文中,提出一種新的方法去分析方筒鏡影像進而偵測出定義這九個視場間關聯性的邊界。在分析的過程中,我們先把影像轉化成灰階來簡化問題,進行邊界偵測的梯度運算後,將梯度運算的結果沿著水平和垂直方向設定一個區域,利用線段濾波器將區域內的影像邊界特徵抽取出來,並以一階線性擬合的方式定義出影像邊界,最後得到定義方筒鏡影像視場間關聯性的邊界點。 為了驗證此一方筒鏡影像分析方法,我們以12個不同背景環境下拍攝的方筒鏡影像進行驗證,經過系統修正以後的平均偵測誤差都在一個令人可以接受的範圍內。
Most of three-dimensional (3D) content generation programs cannot generate a stereoscopic image directly. 3D content is realized by capturing the two-dimensional images of the object at different angles firstly then reconstructing the stereoscopic image by the software. The performance of a stereoscopic image will suffer if the software converting 2D to 3D image is not efficient and accurate. Square tube mirror image system is a special three-dimensional content generation designed to generate a special two-dimensional image composed of nine views (as if taken from nine different cameras), called STM image. The key information to reconstruct 3D image like depth can be extracted from these nine views after accurately segmented from the STM image. In this thesis, we present a method to analyze the STM image by recognizing the boundaries of those nine views. In the process, we simplify the problem by converting to grey scales, confine the edge detection within a certain region of the vertical and horizontal directions by taking x and y gradients of the image, utilize line segment filters to extract the boundaries of STM image, and identify those boundaries by linearly fitting the filtered image. The boundaries are defined, finally, by the four edge points and ready to be segmented. The boundary-recognition results are demonstrated by 12 STM images under different backgrounds. The average detection error after systematic correction is satisfactory.