街景導覽之招牌偵測系統

Abstract

在這份論文中，我們提出了一個以選取邊為基礎的方法來偵測招牌邊緣，我們希望能在街景導覽系統中自動的偵測招牌並在招牌上加入超連結至該店的首頁。整個過程將分成兩個步驟：選取平行候選邊及選取垂直候選邊。在第一個部分中，影像中強度較強的平行邊將會被選取出來，接著因招牌內部文字所產生的文字平行邊將會被刪除；在第二個步驟中，垂直候選邊將從每兩條平行候選邊中選取強度較強的垂直邊，然後因單一文字所產生的文字垂直邊將會被刪除。在刪除平行邊的部分，因為招牌內部的文字與文字之間會有間隔，此種現象在直方圖中會顯示出許多山峰與山谷；在刪除垂直邊的部分，由於招牌文字必定在招牌內部，因此由文字所產生的垂直邊必然會小於招牌所產生的垂直邊。上述特徵皆可以直方圖顯現的平滑曲線所分別。由於招牌通常是大且顯而易見的，因此我們在最後以最大面積選擇法來選出最大面積區域來當作我們的招牌結果。根據實驗結果，我們的方法是有效且快速的。

In this thesis, we present an edge-based shop sign detection algorithm for automatic detecting shop signs in route panorama system. The whole processes are divided into two parts: candidate horizontal line selection and candidate vertical line selection. In the first part, the stronger horizontal lines in the image are selected and then the erroneous lines caused by texts on the shop sign are deleted. In the second part, the stronger vertical lines of the region between two selected horizontal lines are selected and then erroneous lines caused by single character on the shop sign are deleted. For horizontal line deletion, the texts on the shop sign have blanks between each character; this feature generates many peaks and valleys in histogram projection profiles. For vertical lines deletion, characters are on the shop sign and therefore the vertical lines caused by single character are shorter than the vertical lines of the shop sign edges. In contrast, the shop sign edges have a smooth curve on projection profile; this difference can be used to separate them. Since shop signs usually are significant and large in the image, we use the maximum area selection to select the largest region as the shop sign. The experimental results show our method is fast and effective.