一種高維度大型物體資料庫的快速影像搜尋系統

Abstract

本研究計畫是針對數量龐大而又是高維度的物體資料庫進行快速以單張物 體影像就可以捜查到物體的正確名稱及其連貫一起的資訊。3D物體的辨認 能靠一個pose的影像執行，其主要方法有兩種，一種是靠龐大的豐富資料 庫，其內含有該3D物體相近pose的影像所表達模型資料而加以捜尋，另 一種是靠有該物體的3D模型(包括3D形狀模型及3D材質模型）而加以套 合(fitting)與比對。前一方法需要有龐大的資料庫收集物體影像與內容，處 理可快，後者是靠精確3D模型及多張物體照片，處理費時。我們研究目的 是快速物體影像捜尋，故本研究屬於前一型方法，主要研究內容是影像的物 體切割，物體影像內容的表達，以及快速的資料庫1-NN捜尋。我們要針對 這些問題發展有效新技術加以解決。

The goal of this research project is to develop fast content based image retrieval from a high dimensional database using a single object image. There are two major approaches for solving this problem. The first approach depends on the image database containing an object image associated with a pose close to the one of the probe object. The second approach relies on the 3D model of the probe object consisting of a 3D shape model and a 3D texture model. The former needs a huge volume of data space and the latter requires an accurate 3D geometric model and multiple object images. Our main research target is to design a fast search system, so we adopt the first approach, namely, database oriented. Our research subjects include object segmentation from the scene, object content representation, and a speedy 1-NN database search method. We shall solve these issues with new efficient techniques.