現有的圖像檢索系統,除了透過CBIR以圖找圖的方式,找到特徵相近的圖像外,也可使用關鍵字搜尋圖像,如:Google Image Search。若想由圖像找出更多相關資訊,往往需要以該圖像所描述的事物,作為關鍵字,從搜尋引擎尋找到更多相關的資訊;但是人們對於圖像的內容常常很難找出適當的關鍵字以及文字描述,造成人們必須花費許多時間嘗試搜尋甚至無法找到相關資訊。因此本研究提出一套CBIR的方法讓使用者有效利用圖像內容比對,以取得相似圖像之圖像標註,並進一步利用其中的文字描述。 在本研究的架構上,先將資料庫中的圖像利用離散小波轉換藏入圖像的相關文字資訊,再對圖像進行特徵的抽取,包括顏色比例、SIFT特徵描述,以供日後進行相似圖像內容之比對。根據實驗結果顯示,本研究所提方法能準確的找到相似度高的圖像,並且能夠抽取出圖像相關的文字資訊,提供使用者作為進一步搜尋的關鍵字。
In existing image search systems, users can find images that have similar features through content-based image retrieval (CBIR). They can also use the keyword search for images, such as: Google Image Search. In order to find more related images, users often need to provide descriptions of the image as the keywords for search engine to find more relevant information. But it is difficult to find appropriate keywords and text description from the content of the image. It takes a lot of time trying to search from search engine to find relevant information. Therefore, we propose a CBIR system which effectively compare content of image, and obtain similar images and the image annotation embedded in the image. The propose architecture of this paper is as follows. First, we use discrete wavelet transform to hide the relevant text information into the image in database. Then, we extract color ratio and SIFT features descriptors as the image features for similarity matching. The experimental results showed that our proposed approach can accurately find similar images, and extract image-related text information to provide user keywords in search engine.