影像中建築物偵測之研究

Abstract

影像中偵測建築物在影像語意分析上是一個重要的任務。影像語意分析能夠利用一些低階的特徵例如：顏色、紋理、形狀..等來表示，在影像檢索與影像瀏覽上是一項有用的技術，而建築物就是其中一種重要的語意。本篇論文中提出一個以紋理特徵為基礎的建築物偵測方式。方法首先會使用較小區塊中的紋理特徵找出影像中有可能為建築物的區域，再利用較大範圍區域中之紋理特徵來判斷是否為建築物，來找出建築物的區域，研究並延伸至建築物影像分類與特定類型建築物偵測。我以科立爾影像資料庫中的建築物影像來做實驗，測試在本方法下建築物區域偵測結果的召回率與正確率，召回率方面最高能達到75.70％，而正確率可以最高能達到80.32％，結果顯示紋理特徵在建築物偵測有不錯的效能。

Building detection in images is an important task in image semantic analysis. The image semantic analysis can be inferred from lower-level features such as color, texture, shape, exc. which is a very useful technique for the image retrieval and browsing. And building is one kind of the import image semantic. In this paper, we propose a method for building detection based on texture feature. First, we used the texture feature in smaller blocks to find the candidate regions, and then used texture feature in bigger range of regions to determine the correct building regions. The study also extend to building image classification and specific category building detection. The experiments are conducted on the Corel image database. We tested the region’s recall and precision under our method. The best recall rate is 75.70% and the best precision rate is 80.32%. The experimental results show that texture feature is efficient in the detection of building.