| Abstract: | 對大量的藝術與繪畫作品做分類是一件非常繁瑣的工作，而且對繪畫類別的認定也是比較主觀的，本論文提出一個以AdaBoost為訓練方式的繪畫主題搜尋系統，藉由提供一些含有共同繪畫主題的查詢影像，例如肖像畫，並配合回饋的機制，讓使用者可以獲得他想要的繪畫作品。
自從1990年以來，已經有許多的研究在探討如何創造一個理想的模組來描述一張影像的內容，到目前為止的研究發現，以內容為基礎的影像擷取系統（Contented-Based Image Retrieval）所面臨的主要問題是如何有效的消除存在於低階特徵與高階感知之間的鴻溝。為了找出哪些特徵包含著相似的語意，哪些特徵又可以區分不同的語意，本系統利用了一組不小的特徵集合（4,356個特徵），其中包含了SAD、LEP與OC三種特徵。由於繪畫作品不同於自然影像，同樣主題的畫作可能會呈現出全然不同的色調，所以這三種特徵主要是一些繪畫的紋理與空間上的分布組合關係。在使用者提供了初始查詢後，系統利用AdaBoost Algorithm來從這4,356個特徵中挑選出最關鍵的32個特徵來組合成一個分類器來分類資料庫中的影像。雖然AdaBoost可以很有效率的組合出最後的分類器，但由於一個線上的即時查詢系統並不適合要求使用者提供很多的訓練樣本，所以在此提出一個適合本系統的回饋機制，並將AdaBoost做一些修改以期能更有效的利用使用者的回饋來提升分類的正確率。
實驗的結果顯示在一個有634張繪畫的資料庫裡查詢「人物畫」時，在經過3次的回饋之後，其效能可以到達Precision rate = 0.71、Recall rate = 0.84、Top 100 precision rate = 0.95。
A content-based painting image retrieval system based on AdaBoost is proposed. By providing query examples which share the same semantic concepts, e.g., portraits, and incorporating with relevance feedback (RF), the user can acquire painting images he desires.
Despite the great deal of research work dedicated to the exploration of an ideal descriptor for image content since the early 1990’s, content based image retrieval (CBIR) still is crippled from the well known gap between visual features and semantic concepts. In order to find features shared by similar semantic concepts in painting images and able to distinct dissimilar ones, we propose a large set of 4,356 features (including 3 kinds of features: SAD, LEP, OC) based on texture and spatial arrangement of the painting images. After initial query examples and up to three times of RF, the most critical 32 features out of 4,356 are selected by AdaBoost learning algorithm and form a final classifier. We make use of the characteristic of AdaBoost algorithm that it is very efficient in finding a combination of partial weak classifiers, i.e. features, into a strong one, and thus AdaBoost is very suitable for on-line learning. Experiments show a very satisfying result. In query of “portrait” with three RFs, the system shows an approximate 0.71, 0.84, and 0.95 in Precision, Recall, and Top 100 Precision rates. We will try more combinations of features and apply to a larger data base in the near future. |
