用多實例演算法做影像分類與檢索

Abstract

為了搜尋龐大的影像與視訊資料，我們利用多實例學習的架構。 使用者所標記的是一個實例袋而非單一的實例，每個實例袋裡 有多個實例。在訓練資料中， 如果一個實例袋中有一個實例是正面的，則標記該實例袋為正面的； 如果一個實例袋中所有的實例都是負面的，則標記該實例袋為負面的。 我們要訓練出一個可以分類的概念。 我們使用互異密度當做多實例學習的指標，使用EM演算法最大化互異密度。 把這些技巧應用在影像檢索上，每一個影像都是一個實例袋，使用者指定好 正面的實例袋與負面的實例袋後，可以找到類似的影像。

In order to search image and video data, we use a framework called multiple-instance learning. Each example labled over internet by teacher is a bag, consisting of any number of instances. A bag is labeled negative if all instances in it are negative. A bag is labeled positive if at least one of instances in it is positive. We would like to learn a concept which will correctly classify unseen bags. We used Diverse Density algorithm for learning concepts from multiple-instance examples. Then the EM algorithm is used to maximize Diverse Density. Then we applied these techniques to problems in image or video database retrieval. Each image or a video frame is a bag, our system uses a small set of user-selected positive and negative examples to learn a scene concept which is used to retrieve similar images from the database. Based on the similar images, a user can decide the related video.