自動圖像註解於圖像檢索系統之研究

Abstract

圖像自動產生註解文字是一個極具挑戰性且困難的工作，因為其必須找出圖像內容與語意間對應的關係。本篇論文提出一個以圖像區塊為基礎產生自動圖像註解的方式。首先對圖像進行圖像分割找出其可能的物件區塊，並對圖像區塊取出特徵，以此特徵在資料庫中與其他帶有標記字詞資訊的圖像區塊做視覺相似度計算，再將與其視覺相似度較大的圖像區塊所帶的標記字詞設定到該圖像區塊上，並依據平均相似度產生各字詞所對應的權重，最後綜合圖像中各圖像區塊的標記字詞及權重資訊產生圖像註解字詞。我們以Corel圖像資料庫的圖像進行自動圖像註解實驗，驗證所提方法的可行性與正確率，實驗顯示當系統給出1個自動註解字詞時，正確率可以達到40%左右，而在給定15個自動註解字詞時，註解字詞的召回率可以達到50%左右。論文最後更進一步將所提的方法整合至一個既有的以內容為基礎的圖像檢索系統中，提供一個多模式的圖像搜尋與檢索機制，並融入web 2.0的概念，讓圖像提供者也能為圖像區塊加入人工標記字詞，以增進系統的自動圖像註解能力。

Automatic image annotation is a very challenging and difficult task, because it needs to find the relation between image content and semantic words. This paper presents a method for automatic image annotation based on region information of an image. First, images are segmented into regions by their low-level visual features. Second, visual features are extracted from regions and compute the visual similarity with other regions which have tags information in database. Then, tags are assigned to regions which have higher similarity. And weight of each tag is computed. Finally, we merge the information from regions of an image to generate the annotation words. In the experiment, we use the Corel images to verify the accuracy and feasibility of the proposed method. Moreover, the proposed method is integrated into an existed content-based image retrieval system to provide a multimodal approach for image query. We add the web 2.0 concept to make the image provider be able to add tags to regions of the image. By this way, we can maintain or improve the system ability of automatic image annotation.