基於機器學習之飛蝨偵測

Abstract

飛蝨為全球性之重要害蟲，其破壞除了直接吸取作物汁液造成損傷外，部分物種可以傳播水稻縞葉枯病。在這篇論文裡，我們建立一套飛蝨害蟲的資料庫，並且透過機器學習技術，研究開發飛蝨害蟲之分類器技術，藉此提高飛蝨害蟲影像辨識率。完整的流程包含了四個主要部分：影像前處理、候選者偵測、候選者特徵擷取以及分類，其中在分類的部分，我們測試並比較多種分類器，例如：支援向量機（Support Vector Machine，簡稱為SVM）、卷積神經網路( Convolutional Neural Network，簡稱為CNN)等。在我們的實驗中，由於飛蝨的體型較小，以及其體色易受到日照影響而造成在不同個體上呈現不同的顏色，甚至會與背景植株較枯黃的部分混淆，這使我們在辨識上難度增加，因此針對這個問題，我們多使用了自製的特徵擷取方式，以蟲身邊緣與背景植株的強烈差異作為一輔助特徵來強化分類的準確度。實驗結果顯示，我們所使用的特徵與模型能夠提供有效並實用的效果。

Rice Planthoppers are worldwide pests. They can damage plants by drawing juice, and also spread rice stripe blight. In this thesis, we build a rice planthopper data set and develop the classifier using machine learning techniques in order to improve the rice planthopper de-tection rate. The system consists of four main modules: image pre-processing, candidate region detection, candidate region feature extraction, and classification. In the classification module, we test and compare with multiple classifiers, such as Support Vector Machine, Neural Net-work, Adaboost cascade classifier, and Convolution Neural Network. In our experiments, the planthopper’s size is too small, and its color is easily affected by sunshine. This makes differ-ent rice planthoppers different colors. Some of them even look similar to the background. This makes detection much more difficult. To deal with this problem, we create our own feature extraction method, using the edge difference between rice plathopper and background to im-prove the classify rate. Simulation results demonstrate the efficiency and effectiveness of our proposed method.