基於電腦視覺於有害雜草種子辨識系統之研究

Abstract

在植物的進出口貿易中，若當中夾雜著有害外來種植物，對於本國的農業經濟與自然生態是一大危害。其中雜草的入侵途徑，以種子的型態入侵為大宗，且現今在辨識雜草種子工作上，皆以專家使用顯微鏡觀察來進行辨識，是一項耗費眼力與精神的工作，故一套辨識有害雜草種子系統將對防疫工作有所助益。 而若要建構出一個對防疫有幫助的有害雜草種子辨識系統，勢必會遇到有害種子數量稀少難以取得，造成訓練資料不平衡的問題。 在此，本論文的貢獻有兩點：第一，本論文提出一個基於過取樣技術、ISODATA與PCA-LDA轉換之分類器建構方法，並對種子影像抽取出包含大小形狀、顏色、輪廓等28個特徵，以實驗驗證其在訓練資料不平衡的條件環境下，對於辨識有害種子具有不錯的辨識效果。第二，本論文完成一個以電腦視覺為基礎的辨識有害雜草種子系統雛型，為種子辨識專家減輕負擔，期望能提升檢疫防疫的效率，進而加速農產品進出口速度。

Due to the damage of the nature and agricultural economy, to prevent the alien pest plant from the trade practices is very important, and seeds are the most common type for the weed plant invasion. In the recent years, this job was done by human, and it spent a lot of time and harmed human eyesight, so a weed seed recognition system is helpful for the plant quarantine. To build a pest weed seed recognition system, there is a requirements: To build the system by imbalanced training data. There are two contribution of this thesis. First, it propose a method of constructing a classifier based on over-sampling technique, ISODATA and PCA-LDA transformation. And in our experiment, the system worked well to meet the requirement mentioned in the second paragraph. Second, this thesis finished a prototype of pest weed seed recognition to reduce the pressure on the seed experts.