在系統實作的過程中，我們利用圖像資料庫來取得實驗用之圖像，並從中選擇想要建立模組的圖像類型。利用大量的訓練資料，先依照其低階特徵分成若干個群組，再透過特徵擷取子系統分別擷取顏色、形狀、材質等低階特徵，利用兩兩比對的方式可以得到此三種低階特徵的相似度，可以求得對於此物件圖像而言，最佳的一個低階特徵組合，也就是建立此物件種類的比對模組。經過這些模組的建立之後，當使用者輸入一圖像來做檢索時，就可以與這些已建立好的模組作比對，將符合的模組名稱以及模組裡面的圖像回傳給使用者。透過這樣的機制，不僅可以增加圖像檢索的準確率，也可以自動的對圖像加註或是做圖像的分類。 With the development of the internet, the image retrieval systems are very widespread. So the most important thing is to design an image retrieval system with high efficiency and high precision. If we want to design a content-based image retrieval system, there could be a problem that which low level feature is better for comparison. It is suitable for some kinds of object, but the others are not.
The problem could be solved with our system. We adjust the proportion of the low level features by utilizing the similarity of numerous training data of the object. It makes the recognition the object reach highest and there would be a combination of the low level features which is the most suitable with different kinds of object. After the result of the experiment, we establish the associations between image semantics and low level features.
To test the actually system, we utilize the image database to get the image and choose several kinds of object that the module we want to establish. By utilizing a large number of training data, first we classify them into several groups with their low level features, including color, shape and texture. Then we extract the low level features in the feature extraction subsystem. By comparing to each other, we can get these three similarities of the low level features. And then we can find the most suitable combination of the low level features of the object out. That is to say, we establish the comparing modules of the object. After establishing these comparing modules, a user give a query by inputting an image, it will be compared by all the modules we establish. Finally, the name and the images of module will be returned to the user. Through this mechanism, not only the precision of the image retrieval system will be increased but also we can realize image automatic annotation or classification.
