植基於葉片特徵之影像檢索技術之研究

A Study on Image Retrieval Technique Based on the Features of Leaves

張翠雲(Tsui-Yun Chang)

指導教授 ： 黃國峰

國立臺中科技大學/設計學群/多媒體設計系碩士班/碩士(2006年)

摘要

電腦科技的發展，在短短的數十年內，已經使得許多原先看似不可能達成的事情變得具有很高的可行性。藉由電腦的輔助，可以在短時間內處理相當複雜的運算，例如：輸入圖形從資料庫中找尋相似圖形的影像檢索技術。其中利用電腦進行植物類別的辨識，亦是一個值得研究的議題。植物葉片的外型，提供了識別植物的重要資訊。這些資訊主要係來自形狀特徵的描述，例如：緊湊度、細長比等。然而，有許多的特徵描述並無法有效提供正確的形狀資訊，如此將導致影像資料庫搜尋時執行效率較為不彰。本研究將針對植物葉片之形狀特徵開發出一套影像檢索實驗系統，其中採用了對形狀變化較為敏感的系數，針對葉片各項特徵加以檢測。所使用的形狀系數包含有：無單位形狀系數 (Dimensionless shape factor) 與一般形狀系數 (General shape factor) 兩大類特徵值。另外，本論文亦提出了幾項新的形狀描述特徵值。
　　在本研究提出的第一項實驗中，針對葉片影像資料庫，透過三項形狀特徵值進行影像檢索。所使用的特徵值包含有：緊湊度 (Compactness)、細長比 (Elongation) 及對稱性 (Symmetry)。其中，緊湊度與細長比係屬於無單位形狀系數類特徵，然而因為細長比無法有效表達影像像素之分佈狀況。因此本研究納入了一個水平投影的概念，將系數值加以改進，並且提出了一項新的葉片特徵值，稱之為「對稱性」。本實驗資料庫收集了67種值物共670張葉片影像。由實驗結果發現，對稱性確實具偵測出葉片特徵之效用。
　　本研究之第二項實驗主要係進一步利用基礎形狀系數，針對28種理想化全葉形葉片加以分辨。由於，無單位形狀系數無法辨識葉片生長方向的相異性，因此以往的相關實驗，只能有效判別半數 (57%) 的理想化全葉葉片。本實驗針對全葉形植物樣本，進行傳統特徵值的適用性分析。由觀察結果發現，傳統的無單位系數難以分類的對象為：(1) 形狀相同，但互為上下顛倒的兩種葉形、(2) 外型相似，但彼此具有些微差異的葉形。針對此類問題，本研究以重心的概念，發展出兩種新的特徵描述值，包含：T軸比 (T-axis ratio)、面積與高度乘積 (Product of area and height, PAH)。由實驗結果發現，本研究能有效改善傳統方法兩成以上 (21%) 的葉形辨識率。

關鍵字

無單位形狀系數 ； 一般形狀系數 ； 理想化葉片類別 ； 影像資料庫 ； 葉片影像檢索技術

並列摘要

The development of computer technology has made possible many things that seemed almost inconceivable only a few decades ago. By employing computer systems, complex calculations can be processed within a short period of time, such as the retrieval of a digital image from a large database that matches the input image. In this manner, the use of image retrieval systems in identifying different species of plants has become a worthwhile research topic. The leaf shape can provide important information in the identification of plants. The patterns of a leaf can be described by several shape factors, such as compactness and elongation. However, many descriptors are incapable of providing sensitive information about the leaf shape, leading to less efficient system performance while searching an image database. Our research uses dimensionless shape factors and general shape factors to develop a leaf image retrieval system. In addition, several novel shape descriptors are proposed for the identification of idealized leaf types.
　　Using a database of live leaf images, the first scheme developed a leaf image retrieval system based on three shape factors: compactness, elongation and symmetry. Among these factors, compactness and elongation are dimensionless shape descriptors used frequently in many related researches. However, the measurement of elongation is limited in usefulness, since elongation cannot effectively express the distribution of image pixels. Therefore, this study incorporates the concept of horizontal projection to provide more information on pixel distribution in an image. The horizontal symmetric parts of a leaf can be detected using the symmetry descriptor. Experimental results based on 670 leaf images from 67 plants show that the proposed approach can achieve an efficient retrieval performance.
　　The second scheme utilizes general descriptors to identify 28 idealized leaf and petal types. Because dimensionless shape factors cannot be used to recognize leaf types with certain orientations, thus they are only capable of identifying approximately half (57%) of the 28 idealized leaf types. According to the results of this analysis, we found that there are a number of issues that cannot be resolved with traditional dimensionless shape factors, including: (i) differences between two leaves that have the same shape, but with inversed directions, and (ii) similar leave shapes with minor variations. To solve these issues, we presented two new shape descriptors, T-axis ratio and Product of Area and Height (PAH). Results from the experiment incorporating these shape descriptors show a 21% improvement over the traditional method.

並列關鍵字

Dimensionless shape factor ； General shape factor ； Idealized leaf types ； Image database ； Leaf image retrieval technique

參考文獻

[3] Chi, Y.-T., Chien, C.-F., and Lin, T.-T., "Leaf shape modeling and analysis using geometric descriptors derived from bezier curves," Transactions of the ASAE, vol. 45 (6), pp. 175-185, 2002.

[4] Chien, C.-F. and Lin, T.-T., "Leaf area measurement of selected vegetable seedlings using elliptical hough transform," Transactions of the ASAE, vol. 45 (5), pp. 1669-1677, 2002.

[5] Costa, L. d. F. D. and Jr., R. M. C., Shape analysis and classification: Theory and practice: CRC Press, Inc., 2000.

[6] Duda, R. O. and Hart, P., Pattern classification and scene analysis: Wiley, 1973.

[7] Franz, E., Gebhardt, M. R., and Unklesbay, K. B., "The use of local spectral properties of leaves as an aid for identifying weed seedlings in digital images," Transactions of the ASAE, vol. 34 (2), pp. 682-687, 1991.