運用類神經網路進行植物病害影像檢測—以蘭花為例

台灣蘭花出口佔台灣花卉出口總值的92%且2017年外銷金額為1億3,986萬美元，所以台灣蘭花在各類花卉植物中屬於價值較高的品種。因此在蘭花的栽培照顧上是非常重要的，如果能提早發覺蘭花是否發生異常，可以避免造成更多傳染並提早治療。過去學者也做過很多植物影像病害的辨識，例如：香蕉葉、苜蓿葉、柑橘葉等等許多植物，透過影像處理技術對影像進行簡化分析找出病害特徵辨識出植物的健康狀況。然而，過去以單一植物的特徵研究為主，對於種植多種植物的農民在照顧上較缺乏相關的方法，也比較少文獻提到透過影像辨識檢視協助花農大面積的遠距離監控培養。因此如何運用影像做遠距離的植物照護，了解當前植物狀況與辨識多種植物的成長情形，對現代的花農而言將是個很重要的任務。基於上述的問題，本研究提出結合影像處理技術與類神經網路來建立類神經網路模型來進行蘭花的病害辨識。透過蘭花葉子影像之HSI、RGB及灰階色彩等特徵空間資料，再運用直方圖分析找出葉子顏色(綠色)像素遮蔽的閥值，運用此閥值對葉子影像進行分割來找出病害區塊，最後使用類神經網路(Artificial Neural Network)對蘭花的葉子進行辨識，並判斷蘭花是否為健康或生病的狀態。本研究比較了人工神經網路、卷積神經網路、深度神經網路等機器學習模型，利用輸入資料的不同特徵和網路的參數設定來找出這三種模型對於蘭花病徵判斷的結果及影響。

關鍵字

類神經網路；植物病害；影像檢測；蘭花；深度學習

並列摘要

The orchid exports account for 92% of the total value of Taiwan's flower exports and the export value in 2017 was the US $ 139.86 million, so Taiwan orchids are among the higher value varieties of various flower plants. Therefore, it is very important to cultivate and care for orchids. If you can detect whether orchids are abnormal, you can avoid causing more infections and treat them early. In the past, scholars have also done a lot of plant image disease identification, such as banana leaves, alfalfa leaves, citrus leaves, and many other plants. Through image processing technology, a simplified analysis of the images is performed to find out the disease characteristics and identify the plant health status. However, in the past, the study of the characteristics of a single plant was mainly used, and there was a lack of relevant methods for the care of farmers who planted a variety of plants. There are also few references in the literature to assist flower farmers in large-scale remote monitoring and cultivation through image recognition and inspection. Therefore, how to use images to do long-term plant care, understand the current plant status and identify the growth of a variety of plants will be an important task for modern flower farmers. Based on the above problems, this study proposes to combine image processing technology and neural network to build a neural network model to identify orchid diseases. Through the feature space data of orchid leaf image HSI, RGB and grayscale color, and then use histogram analysis to find the threshold value of leaf color (green) pixel masking, use this threshold to segment the leaf image to find disease blocks , And finally use the Artificial Neural Network to identify the leaves of the orchid and determine whether the orchid is healthy or sick. This study compares machine learning models such as artificial neural networks, convolutional neural networks, and deep neural networks, and uses different features of input data and network parameter settings to find the results and impact of these three models on the diagnosis of orchid disease.