Our system overview We input a photo to our system Result Input Photo Result Input photo Result
Building Recognition from Architectural Photos Collection 〔開發由建築物相片集中辨識特定樓房之技術〕

In recent years, web services, such as ‘’Yahoo Image Search,’’ ‘’TinyEye,’’ ‘’ookaboo,’’ that provide image search engine have become popular. Benefiting from these image search engines, users can input their photos to find the pictures that contain similar object, therefore users can recognize the contents in the photos they uploaded. If the uploaded images happened to be building photos, we can develop a tourism system providing information for users. However, searching the building photos accurately is a challenge, due to the variation of lighting conditions, diverse viewpoints and varieties of textures, which affects the photo quality and searching results. In this thesis, we proposed an image search system that aims at searching building and landmark architectural photos. With taking at least 1024x768 resolution of color building image as an input, our system can find photos that contain similar or the same buildings from our database. We exploited a three-step method which combines color feature, Histogram of Oriented Gradient and Scale invariant feature transform to solve the problems. To improve the performance of searching speed, we use tree structure and string metric method to make our system more efficient.

SUMMARY (中文總結):
近年來,愈來愈多網站像是『谷歌』、『雅虎』、『台妮艾』提供圖像搜尋的功能。 藉由這些圖像搜尋引擎的幫助,人們可以上傳照片加以搜索並找到含有相同或類似物件的照片,因此使用者可以了解他們所上傳的照片內的東西為何。 如果上傳的照片專一於建築物照片,那麼我們即可以利用這項技術發展導覽系統提供給使用者們。 然而,如何準確搜尋建築物照片是一個挑戰,由於光線的變化、多變的視角和不同的紋理材質都會影響到照片的品質和搜尋結果。 輸入一張解晰度為1024x768的彩色建築物照片,我們的系統會自動比對存於資料庫的照片並回傳含有相同建築物之照片。 我們的方法分成三個步驟分別利用色彩特徵、梯度直方圖及尺度不變特徵轉換以解決光線的變化、多變的視角和不同的紋理材質的問題。 為了改善系統效能,我們利用樹狀資料結構及字串比對方法加速。

首先我們會將存在資料庫的照片的特徵取出,主要有顏色特徵、HoG及Visual word,並將這些特徵預先存在資料庫內,同時也利用kd-tree資料結構 組合visual word特徵,以加速之後比對的速度。
首先當使用者輸入一張欲搜尋的照片,我們會先請使用者圈選照片內的建築物是否有較明顯的幾合結構特徵,如果有的話,我們會根據圈選出來的 部份進行比對,否則,我們皆以整張照片進行比對,接著我們的系統設計,依顏色、HoG、visual word的順序進行比對,以找出相似的照片。