稻作蟲害監測和預警系統

Abstract

飛蝨為一種全球性農業稻作害蟲，隨著強風由菲律賓大批遷徙至台灣，飛蝨會直接吸取作物汁液，造成稻米嚴重的損害，此外還會傳播水稻疾病，使稻米產量大減，因此如何有效的撲殺飛蝨成為了非常重要的議題。 在這篇論文中，我們提出了一個飛蝨害蟲的監測系統，利用影像辨識的技術計算當下飛蝨的數量，當數量到達一定指標才會噴灑農藥，能夠最佳化噴藥的效益，也能避免過度的噴藥對土地造成影響，監測系統會分為兩個階段，第一階段會先使用傳統的影像處理技術找出植株位置，切掉周圍非植株部分供第二階段辨識，第二階段使用了深度捲積神經網路偵測飛蝨的位置，此階段我們改良原先Single Shot Multi-box Detector(SSD) 的模型，原先SSD的架構使用了最大池化層(Max Pooling Layer)降低特徵圖的大小，最大池化層能大幅降低參數量，但也因此流失了許多背景資訊，導致分類時常常把背景的光源判斷成飛蝨，為了解決此問題，我們設計了一種新的池化方式，借由保留區域像數差值的方式輔助原有的SSD架構，此架構更能保留背景的資訊，因而在分辨率上有顯著的提升。

Planthopper is a kind of rice pest that travels from Philippines to Taiwan every year. These pests are able to polish off rice in a short time and spread quickly to a large area. Moreover, they carry some sort of disease that makes rice sick. When they travel from one place to another, the disease spreads along these areas and cause huge damage to the agricultural industry. For these reasons, it is urgent to construct a system that is capable of detecting rice pests in time. Our detection model consists of two stages. At the first stage, traditional image processing technique conducts to detect the main part of the plant. We reserve this part and discard the remaining of the image. At the second stage, we use the convolutional neural network to detect pests. The model is based on the Single Shot Multi-box Detector which performs classification and localization at the same time. However, SSD has a serious problem for discarding too much background information at the max pooling layers. It is easy for the model to misrecognize reflected light as positive due to their similar shape and color. To solve this problem, we introduce a new way of pooling. Instead of reserving the max value, our model intends to save the local difference. This model can reserve more background information and obtain better results when evaluating.