- 學位論文
使用者介入之物體照片查詢系統-以台灣常見鳥為例
An User Augmented Object Query System for Photos of Common Wild Birds in Taiwan
摘要
本研究嘗試提出一套基於使用者介入的色彩及形狀特徵的影像查詢系統，希望能讓使用者基於圖片及人機互動的方式更快找到物目標物種。 本研究共收集網際網路上的1708張鳥類影像樣本做為訓練集及測試集。在色彩方面，本研究採用可濃縮色彩維度的HMMD色彩空間，搭配色彩結構描述子的頻率作為特徵向量；而特徵值比對方法則使用了歐氏距離及C5.0決策樹做為決策方法；在形狀方面，本研究試著透過使用者介入的方式，取得鳥喙及頭部的長寬比做為特徵值。 在實驗部分，主要將測試對象分為四部分進行，(1)以台灣常見的98種鳥種為實驗對象；(2)以台灣常見的98種鳥種進行K-Means分群後進行實驗；(3)以10個不同棲息地中的常見鳥為實驗對象；(4)以22個不同科別的常見鳥為實驗對象，針對上述的方法分別以歐氏距離及決策樹為基礎，測試其正確率。 實驗結果顯示，未將鳥種分類時的精確度並不夠理想(約27%)，若能夠知道該鳥種的棲息地或科別等資訊，則可取得精確度較高的預測結果(決策樹約80%)。在未來的研究方面可透過增加紋理的特徵值嘗試進行系統改善。
並列摘要
This paper proposes a user-augmented object query system based on the color and shape feature of Taiwan wild bird photos. The goal is to allow users to search target species faster by photo and human-machine interaction. For each bird photo, the color feature is the 32 dimensional HMMD color structure descriptors and, the shape feature is the beak-width, beak-height and the head diameter of the target bird. Using these features, Euclidean distance and C5.0 decision tree are tested to determine the target species of the query photo. 1,708 photos were collected from the internet as the training and testing corpora. The experiments are divided into four types. (1)The most common 98 bird species; (2) the most common 98 bird species with K-Means Clustering; (3) the birds in ten types of habitat; (4) the birds in 22 families. The Euclidean distance and the decision tree model are graduated to test the correct rate. The experimental results show, that using methods without classification or K-Means Clustering are unsatisfied (about 27%), but if the user can know the habitat or the family-of-kind of the bird, the precision will conspicuous increase (the precision of C5.0 is about 80%).