- 學位論文
以影像為基礎之植物辨識及問答系統
An Image-based Plant Recognition and Question-Answer System
摘要
一般民眾在認識植物的時候，是以圖鑑紙本翻閱查詢，這種方法在查詢上不但很耗費時間，而且圖鑑內植物的種類不一定齊全。本論文提出以植物各部位的影像來自動辨識物種，包括花朵與葉子。不同種類植物的花朵除了顏色有差異，外型輪廓也會不同；葉子通常皆為綠色，但是葉子的輪廓為辨識特徵。我們蒐集50種常見植物的花朵與葉子影像，並利用圓形區域色彩直方圖及一維傅立葉描述子進行顏色及輪廓的辨識實驗。除此之外，本論文以單一部位多種特徵合併排序與多部位多種特徵合併排序的方法，分別以 Linear Evidence Combination、Dempster-Shafer Evidence Combination 及 Plant Information Fusion 進行比對實驗。花朵的多特徵合併方式，Linear Evidence Combination的 Top 10 辨識率有87%。Plant Information Fusion 合併花朵顏色、花朵輪廓及葉子輪廓的 Top 10 辨識率達 96.4%。 然而，植物的物種實在是太多了，使得自動影像辨識的辨識率不高。所以本論文也設計一套能讓使用者詢問植物問題的行動應用程式，藉此來吸引群眾參與認識植物，並且以群眾外包的方式大量地蒐集植物影像。
並列摘要
When the general public want to understand the plants, they use plants illustrated guides. This method is time-consuming, and the kinds of plants in plants illustrated guides are not complete. This thesis proposes to automatically identify species by the images of different parts of plants, including flowers and leaves. Flowers of different plants differ not only in colors, but also in their contours. Leaf color is usually green, but the contours of the leaves can be the identification features. We collect flower and leaf images of 50 common plants, and apply Circular Color Histogram and 1-D Fourier Descriptors as color and contour features in the recognition experiments. In addition, this thesis conducted experiments of multi-features fusion in a single part and multi-features fusion in multiple parts by applying Linear Evidence Combination, Dempster-Shafer Evidence Combination and Plant Information Fusion. Multi-features fusion of flowers reached Top 10 recognition rate 87% using Linear Evidence Combination. Applying Plant Information Fusion, multi-features fusion of flowers’ color and contours and leaf contours reached Top 10 recognition rate 96.4%. However, there are too many species of plants such that the recognition rate is not satisfactory. This thesis also designs a mobile application for question-answering for plants to attract people understanding plants. In this way, we can gather a large number of plant images by crowdsourcing.