影像自動辨識技術在檢防疫介殼蟲害鑑定之應用

Abstract

介殼蟲為台灣水果產業上的害蟲，近年更是造成我國重大的出口損失。傳統上果物表皮是否有介殼蟲係用人工肉眼的方式進行辨識，然而人工檢查耗時費力，且出口量龐大，無法逐顆仔細檢測，導致近年在他國海關檢疫出介殼蟲而導致禁止進口的案例發生，對於大量依賴出口的台灣水果經濟而言，實屬重大危機。為了有效地避免介殼蟲對於水果的風味、賣像以及出口造成影響，我們採取自動化方式，利用機器視覺與深度學習的方式，來達到快速、省力的輔助檢查手段。本研究利用物件偵測卷積神經網路模型——YOLOv5 進行介殼蟲辨識。我們建置三個物件偵測模型分別用以檢測鳳梨、釋迦與檸檬三種果物表面的介殼蟲，以期能幫助實現國內果物出口前的介殼蟲害自動檢測。本研究透過農試所研究人員協助蒐集影像進行模型偵測。經測試得到鳳梨水果之模型準確率（mAP@0.5）粉介殼蟲為0.558 和褐圓盾介殼蟲為0.371、釋迦模型準確率粉介殼蟲為0.8001、檸檬模型準確率黑片圓盾介殼蟲為0.7956。可以看到目前在處理介殼蟲害上釋迦和檸檬的模型已經可以有效率地偵測介殼蟲，但是鳳梨因為表皮紋路複雜且有縫隙不利於蟲體辨識，因此在準確率上的表現不如另外兩種果物。這個研究提出的方法可以提升檢測的效率與品質，避免因人工辨識耗時費力，無法逐⼀檢查的窘境，並提供介殼蟲出口檢測適當的輔助，以維護台灣農產品出口的品質。 Scale insects are pest in Taiwan’s fruit industry. Recently, they have caused great export losses to our country. Traditionally, whether scale insects existing on fruit peels are inspected by naked eye. However, manual inspection is time-consuming and laborious, and due to the immense amount of export, it is impossible to inspect scale insects thoroughly. In recent years, customs of other countries still have quarantined scale insects on fruits. It is a serious crisis for Taiwan’s fruit economy which heavily relies on exports. In order to effectively avoid the impact of flavor, appearance, and export of fruits, we decided to adopt automatic methods, utilizing machine vision and machine deep learning to achieve rapid and labor-saving auxiliary detection methods. This study uses object detection convolutional neural network, YOLOv5, to identify scale insects. To accomplish automatic detection before exporting fruits, we constructed three object detection models to detect scale insects on the peels of pineapple, sweet apples, and lemons. In this research, the expert in Taiwan Agricultural Research Institute Council of Agriculture assisted in image collection for model detection. After testing, the model accuracy(mAP@0.5) of pineapple was 0.558 for scale insect and 0.371 for Chrysomphalus aonidum, the model accuracy of sweet apple was 0.8001 for scale insects, the model accuracy of lemons was 0.7956 for Parlatoria ziziphi. It can be seen that on detecting scale insects, the sweet apple and lemon models are able to detect scale insects effectively, but the pineapple model are not as accurate as the other two due to the complex peel texture and seams which are inconducive for pest identification. The method proposed in this study can improve the efficiency and quality of inspection, and provide appropriate assistance of pest quarantine to maintain the quality of Taiwan’s produce export.