針對電子書閱讀之目的作影像內容索引建構與內容搜尋之研究

Abstract

近年來由於電腦技術日趨成熟，各式儲存裝置的容量大大增加，數位形式的聲音、影像的使用，也相對地被廣泛地使用。大量使用數位影像資料的結果，產生了另一個問題：即如何自動且快速地找到所需的影像。本論文針對此問題，發展出一套適用於電子書平台上自動建構影像索引及作自動影像搜尋的系統。首先我們依影像色彩及灰階的性質將影像分成數類。影像分類可幫助我們建置出影像索引，而且加快影像搜尋的速度。若影像數目眾多，影像搜尋系統可幫助我們快速地找尋到我們所需要的影像。在影像搜尋系統中，我們使用了兩個特徵：色彩及紋理。在色彩方面，我們利用全彩階層式矩量保持原理來做影像減色。做影像減色可加快搜尋的速度及消除雜訊。接著我們利用線性規劃中的運輸單純化方法，來做相似度量測。此單純化方法是用來解最佳化問題的一種方法。兩張影像在減色後，其相似度量測值越高，代表這兩張影像越相似。在紋理方面，我們利用空間灰階相依方法，加上能量量測，成功地判別出影像紋理的方向性。我們利用色彩及紋理特徵擷取整張或部分影像。因為紋理的方向性在小範圍影像中才具意義，所以我們使用色彩特徵擷取整張影像；使用色彩及紋理特徵擷取部分影像。藉由實驗的結果，可以證明我們提出的系統與方法是可行的。

In recent years, owing to the maturity of the computer technology, storage devices of all kinds enlarge their capacity intensely. In consequence, images in digital forms are used more widely than before. Accompanied with the more frequent usage of our digital images, there comes an annoying problem – how we can find the desired image as soon as possible, especially in a large image database. Focusing the problem, a system which can index and retrieve images automatically is developed in this study, which is applicable on the platform of an e-book. First, we classify images into several categories by the features of colors and gray scales. Such image categorization helps us not only construct the image index but also increase the speed of image retrieval. Two main features, color and texture are used in our retrieval system. As for colors, we perform color reduction by a hierarchical moment-preserving method. The process of color reduction improves the retrieval speed and eliminates noise. Then we measure the similarity between two images by the transportation simplex method. The larger the similarity value is, the more alike they are. As for texture, we detect the direction of the image texture successfully by a spatial gray-level dependence method and an energy measure. We can retrieve images entirely and partially by the color and texture features. Because the direction of the texture is meaningful in small parts of images, we retrieve desired images by full sized query images using the color feature, and by partial query images using the color and texture features. Experimental results show the feasibility and practicability of the proposed approaches.