一個用於歪斜復原文件影像的無特徵點接圖法

A Featureless Image Registration Method for Deskewed Document Images

本論文提出一個用於歪斜復原文件影像的無特徵點接圖法，首先對欲接合的文件影像進行歪斜復原，再對這些具有重疊區域的影像進行接合，共分為三個主要程序。在前處理的程序中，我們先將所擷取到的彩色文件影像轉換為二值化影像，然後利用相連元件偵測演算法來過濾文件影像中的非文字資訊。在歪斜復原的程序中，我們採用分析投影圖的方式來偵測文件影像的歪斜角度，它係根據階層式搜尋法的概念，逐步逼近找出一精確的歪斜角度。而在文件影像接合的程序裡，我們發展一個用於歪斜復原影像的無特徵點接圖法，此係以階層式區塊比對的方法，獲得相鄰重疊影像之間的平移參數及縮放參數。截至目前的實驗結果顯示：我們所提的方法能夠正確並且快速地復原歪斜文件影像及接合多張復原後的重疊影像。

In this paper, we present a featureless image registration method for deskewed document images, which first recovers skewed document images and further registers them with overlapped regions. The following are the three main procedures in this document image registration method. In the preprocessing procedure, the original color document image is transformed into a binary one. A connected component detection algorithm is then applied to filtering non-textual information in the document image. In the skew recovery procedure, a projection profile analysis is adopted to detect the skew angle of a document image with a hierarchical search scheme. In the document image registration procedure, we develop a featureless approach to registering overlapped recovered images, which is mainly basedon a hierarchical block matching to acquire translation and scale parameters between two such images. So far, our proposed method can recovery skewed document images and register multiple overlapped recovered images correctly and fluently.