SURF-based Mobile Object Image Recognition and Its Applications

[[abstract]]本研究提出兩種以SURF為影像特徵為基礎之行動影像辨識方法，並分別作用在以飲料瓶罐與房仲業廣告傳單DM為辨識標的物的應用上，使用者使用行動裝置拍攝瓶裝飲料包裝或傳單上的圖片就可得到商品的相關訊息，本論文的研究成果可以延伸應用於行動電子商務與行銷的目的。有別於目前普遍使用的QR Code行動辨識方式，所提出的方法不需要特別加註標記於物品上，直接針對瓶罐包裝與傳單圖片進行辨識，美觀且不必佔據物品版面。 本研究所開發之系統利用Client-Server架構進行物件辨識，Client端為行動裝置，透過行動裝置進行影像擷取並將影像上傳至Server端進行影像處理與辨識。Server端將影像透過SURF取得特徵點，再經由特徵點匹配演算法計算所擷取的影像與資料庫中的影像的匹配情形，匹配點數越多表示兩張影像越相似。但為了系統實作之效能考量，我們針對瓶裝飲料與房仲廣告辨識之應用分別進行不同的前處理。針對瓶裝飲料的應用，為了解決部分飲料包裝常出現不同口味的飲料，其影像內容幾乎相同，只有顏色不同的情況，本研究透過HSV色彩空間轉換將包裝影像轉換成360個色系(hue)的向量直方圖，經過K-means分群成5類，透過分群可以將外觀相似物件但顏色分布差異較大的飲料包裝進行有效區分，除了可以增加辨識率，同時也減少比對數量，加速比對時間。在房仲業廣告傳單部分，則是將影像中的SURF特徵點進行BoF(Bag of features)的特徵計算，透過BoF將特徵點分布相似的影像分成3群以減少比對運算時間。 經由實驗結果顯示，所提出方法在包裝飲料應用的正確辨識率為94%，平均執行時間為1 04秒；在房仲業廣告傳單正確辨識率為100%，平均執行時間為5 21秒，實驗結果說明所提出方法具有高度的穩定性與可行性。[[abstract]]This study proposes two mobile object recognition methods based on SURF and were applied on the applications of beverage bottles and Real Estate DM respectively Users can use a mobile camera to shoot bottled beverage or pictures on Real Estate DM to obtain more information for the target objects The results of the research can be extended and applied on mobile e-commerce and marketing purposes Different to the current widespread use of QR Code the proposed method can provide the same information services without adding extra-marks on the products This developed mobile object recognition system is based on a Client-Server Architecture In the client-side the mobile device will capture target object images and then upload them to the Server-side In the server-side SURF based image feature extraction and image matching is performed The matching results will send back to the client-side In order to obtain a high performance of systems two different preprocessing procedures for bottled drinks and Real Estate advertising were applied In the application of bottled beverages HSV color space transformation and K-means clustering are applied to increase the accuracy recognition rate In the Real Estate DM application BoF (Bag of features) is applied The experimental results show that the proposed method can achieve high accuracy rates The accuracy rate for packaged beverage application using the proposed method is 94% and the average execution time is 1 04 seconds; in the Realty Estate DM application the accuracy rate is 100% and the average execution time is 5 21 seconds Experimental results show that the proposed methods own a high degree of stability and possibilit