利用特徵分層和權重相似性測量建置葉片檢索系統

目前常用於葉片影像檢索的形狀特徵有Zernike moments、長寬比、離心率、緊湊率等，而邊緣特徵則是質心距離。這些特徵都是可以有效描述圖片的區域形狀，透過對於影像的描述能力以及抗旋轉、平移和不變性的特質，可以幫助葉片檢索做更進一步的有效特徵抽取。本研究在葉片檢索系統中提出利用數值分類學的概念來進行特徵分層的設計，透過階層集群分析法得到最佳的特徵分層順序，至於權重則是採用歐幾里德距離與相似性關係進行計算。當歐幾里德距離的值越大，表示相似性越低，所以給予的權重則要越小，反之亦然，透過這樣的計算方式得到權重的計算公式，並與特徵分層一起計算影像之間的相似性程度。最後利用特徵分層的設計方式，與歐幾里德距離與相似性之間的對比關係所提出的權重計算，建立葉片檢索系統的架構，先將抽取出來的葉片特徵量化成特徵值，並進行正規化的計算，再進行特徵分層和權重計算的相似性測量，最後找到相似的葉片影像，並將葉片檢索結果完整呈現給使用者。透過有特徵分層與權重計算的葉片檢索系統，不但能讓使用者快速查詢出葉片的種類，也可以找到想要瞭解的葉片資訊。實驗結果顯示，有使用特徵分層加權和未使用特徵分層加權的方法，還有使用特徵分層但未使用加權的方法，以及葉片分類所提出的長寬比、邊緣比及中心位置的加權方式和長寬比、邊緣比、中心位置及投影的加權方法一起進行績效評估後，本研究所提出的特徵分層加權方法來進行葉片檢索，最後得到的檢索結果的確可以提高其準確性，讓檢索的績效也可以越來越好。

關鍵字

葉片檢索系統；特徵分層；權重；數值分類學；相似性測量

並列摘要

Currently the shape features of the leaf image retrieval of are Zernike moments, aspect ratio, eccentricity, compact rate. The edge feature is the centroid distance. These features can describe the shape of the image area effectively. Ability to describe the use of images and anti-rotation, translation and invariance characteristics, can help leaf retrieval system to make more efficient retrieve feature extraction. Leaf retrieval system uses multiple features to compute similarity, given the weight of each feature are the same, you can’t distinguish the features of importance for leaf retrieval system. In order to use weights to indicate the importance of leaf characteristics, this study proposes stratified features and weighting similarity measure for similar leaf retrieval. Design and combination of stratified features is based on the concept of numerical taxonomy and hierarchical clustering. Weighting similarity measure involves Euclidean distance and a weighting function. When the Euclidean distance the greater the value that the lower the similarity, so the weight will have to give the smaller, and vice versa. Use this formula to get the weighting, and calculate and features layered with the degree of similarity between images. Finally, the use of stratified features design approach and weighting similarity measure involves Euclidean distance and a weighting function in the establishment of leaf retrieval system architecture. First step is to extract the values of features, and the normalization of the calculation. Then stratified features and weighting similarity measure are performed. The proposed system can locate similar leaf images and return search results to the user. Through stratified features and weighting of leaf retrieval system, users not only allows to quickly find out the types of leaves, but also find the information they want to know the leaves. Experimental results show that using the stratified features and weighting method, using the stratified features, but not using weighting methods, the proposed leaf classification to use aspect ratio, edge ratio and the center mode methods and aspect ratio, edge ratio, central location and projection method with the calculate performance evaluation. In this study the proposed method can improve the retrieval performance and finally get the search results can indeed improve their accuracy, so retrieval performance can be better.