A Location and Image based Plant Recognition and Recording System

[[abstract]]台灣擁有多樣的植物生態，而這些植物更是跟我們的生活密切相關，例如: 植物提供我們食用的蔬果、或蝴蝶所採集的花蜜，然而我們對這些植物的認識卻很少，而一般民眾查詢植物種類常使用紙本圖鑑，並且常用科別當作索引，對較不熟悉植物科別的民眾，此查詢方式較無效率。 因為一般民眾辨識植物的困難，於是本論文提出以適地性搜尋引擎 (Location-Based Search Engine) 輔助「以圖找圖」的自動植物葉子辨識。首先研究去背景，本論文採用常見的去背景方法，分別為Otsu法、Color Slicing與GrabCut，利用影像分割率評估方法找出最佳的去背景方法。之後使用一維傅立葉描述子、二維傅立葉描述子等葉子特徵進行植物辨識的相關實驗，再找出最佳的植物辨識的方法設計出一套能植物辨識、調查植物詳細資料(如地理位置、植物特徵描述) 的行動應用程式來吸引群眾參與，以群眾外包的概念進行蒐集大量的植物影像與地理位置資訊。 本論文的植物影像資料庫是以國立暨南國際大學的常見植物種類為主。為了找出最佳的去背景方法與植物特徵，首先進行小規模的植物辨識實驗。在室內環境取像21種植物種類，每一種類有20張影像總共420張，透過實驗分析後，最佳的去背景方法是GrabCut(自訂遮罩)與最佳的植物特徵是一維傅立葉描述子。接著將此研究實作於系統上，在實地環境收集50種植物種類，每一種類有10張總共500張，並透過適地性搜尋引擎來輔助植物辨識。實驗結果的Top1平均準確度高達78%。[[abstract]]Diversity of plants is rich in Taiwan. These plants are close to our life. For example, the fruits and the vegetables are used for food, the nectar is collected by butterflies. However, many plants are unfamiliar to us. People search the plant species by plant illustrated handbooks which use the plant’s family as the index. This query method is inefficient for people who are unfamiliar to plant’s family. Because of the difficulties to identify the plants for people, the thesis proposes to automatically identify the leaves applying content-based image retrieval assisted with Location-Based Search Engine. The first step is background subtraction. We adopt common methods which are the Otsu method, Color Slicing and GrabCut. We find the best method by evaluating image segmentation rate. Then the plant features are extracted using 1-D Fourier descriptors and 2-D Fourier descriptors. We conducted experiments to find effective features to design mobile application which can automatically recognize plant species. The application applies the idea of Crowdsourcing to attract people joining survey of plant information (example: location, description of plant), in order to collection a large number of plant images and geolocation. In this thesis, we collect plant image database based on common plant species in NCNU (National Chi Nan University) .In order to find the best background subtraction and plant features, we conducted a small-scale plant recognition experiment. For 21 plant species, 20 pictures for each plant species were collected at indoor. Through the experimental analysis, we find the best of background subtraction is GrabCut (custom mask) and the best plant features is 1-D Fourier descriptor. These methods are implemented on the mobile application. In the system, we collected 50 plant species and each plant species has 10 pictures taken at outdoor environment. Using Location-Based Search Engine to assist plant recognition, experimental result reached 78% Top 1 average precision.[[note]]碩