隨著數位影像資料庫使用越來越普及,以“內容為基礎之影像檢索”的技術研發成為成為熱門的研究領域。本研究提出一個以隱藏式馬可夫模式 ( Hidden Markov Model ) 為基礎的階層式形狀檢索方法。在本方法中,我們將整體性及細緻性特徵及特徵間的脈絡關係皆巧妙地納入隱藏式馬可夫模式中,使用整體及細緻相似度皆納入考量以提高檢索相似度。另外由於採用的特徵皆具有旋轉、平移及大小比例變化之不變性,而隱藏式馬可夫模式又具有強韌的容錯性及融入脈絡資訊之能力,再加上本法採用由簡至繁的階層式處理方法,故所提方法不但具有旋轉、平移及大小比例變化之不變性,且對於變形如透視、扭曲及遮蔽都具有不錯的強韌性。本法驗證於三種不同形式的資料庫,皆證實能正確地檢索具有相似物件形狀的影像。
With the increasing popularity of the use of digital libraries, it becomes imperative to build an efficient content-based image retrieval system to browse through the entire library. Since the unit of an image is shape, a new shape-based image retrieval method is proposed in this study. Coarse and fine shape features as well as the statistical and the contextual information are incorporated into the hidden Markov model hierarchically to calculate a probability value. Then, the probability values can be considered the matching scores to retrieve similar shapes. In addition to the proposed method being translation, rotation and scaling invariant, it is robust to various distortions such as perspective, shear and occlusion. These advantages are accomplished by the strategies we adopt. First, the proposed features, either coarse or fine features, are translation, rotation and scaling invariant. Secondly, the flexibility of shape comparison can be increased since the HMM is high tolerance to noise and distortion and into which the contextual information can be incorporated. Finally, coarse and fine matching using HMM are hierarchically employed to include overall and fine detailed information both of which are necessary for similar shape ranking. The proposed method has been applied on three databases: geometry, character and map. The experimental results prove the effectiveness, robustness and practicability of the proposed approach.