區域卷積神經網路及位置回歸方法於茶葉病蟲害辨識之應用

茶樹病蟲害是造成茶葉病變的主要成因，且會導致龐大的經濟損失。為了進行適當的防治措施，需要在病蟲害發生早期就辨識其成因。一般而言，茶葉病蟲害須由具備專業知識的人員來進行判斷。然而目前專業人才有限，故無法滿足及時病蟲害辨識之需求。本研究旨在建構一即時辨識系統以達成茶葉病蟲害辨識及其危害程度判別，並收集4295張茶葉表面病蟲害影像，其中包含三種病害及六種蟲害，病害分別為茶赤葉枯病、茶餅病及茶藻斑病等；蟲害分別為潛葉蠅、茶黃薊馬、茶角盲椿象、茶捲葉蛾、茶姬捲葉蛾及黑姬捲葉蛾等。病蟲害影像用於深度學習模型之訓練，而辨識模型採用更快速卷積神經網路(faster region-based convolution neural network, FRCNN)及串接卷積神經網路(cascade R-CNN, CRCNN)。並比較三種不同位置回歸損失函數於模型效能之增進，分別為矩形框偏移差(bounding box difference)、概括交集聯集比(generalize-IoU loss, g-IoU)及距離交集聯集比(distance-IoU loss, d-IoU)。開發之模型效能於使用FRCNN模型時達74.0% 之平均精確度(average precision, AP)及77.9% 之分類準確度(accuracy of classification, Acc)；於使用CRCNN模型時達到75.8% 之平均精確度及83.6% 之分類準確度。本研究建立之模型可於網頁伺服器上提供茶葉病蟲害辨識之服務。藉由推廣此服務可協助茶農即時了解茶園中的病蟲危害，並規劃病蟲害整合管理，以降低病蟲害造成之茶產量損失。

關鍵字

茶葉病蟲害；病害辨識；位置回歸損失函數；串接卷積神經網路

並列摘要

Tea diseases and harming insects are major factors that cause lesions to tea foilages and result tremendous economic losses. It is essential to identify the lesion causes in a relatively early stage so that proper actions can be taken. Conventionally lesion causes were identified manually by phytopathologists or entomopathogists. However, professionals are limited and may not meet the demands of lesion cause identification in time. Thus, an automatic identification approach is needed. This study aimed at developing a real-time identification system for identifying the causes of lesions and the severity stages of the lesions for tea foilages. A dataset composed of 4295 tea foilage images with lesions were collected. The causes of the lesions included three diseases – brown blight (Colletotrichum camelliae), blister blight (Exobasidium vexans Massee), and algal leaf spot (Cephaleuros virescens Kunze); six harming insects – tea mosquito bug (Helopeltis fasciaticollis Poppius), oriental tea tortrix (Homona magnanima Diakonoff), small tea tortrix (Adoxophyes sp.), tea flush worm (Cydia leucostoma Meyrick), tea thrips (Scirtothrips dorsalis), and leaf miner (Liriomyza sp.). A faster region-based convolution neural network (FRCNN) and a cascade RCNN (CRCNN) were then trained to identify the causes of the lesions. Three training losses were used, including bounding box difference, generalize-IoU loss (g-IoU loss), and distance-IoU loss (d-IoU loss). The trained FR-CNN model achieved a mean average precision (mAP) of 74.0% and accuracy of classification of 78.0%. The CRCNN achieved an mAP of 75.8% and an accuracy of 83.6%. The trained models were hosted on a webpage to provide the service of lesion cuase identification to public. The developed system provides real-time and may help farmers set up a better-integrated pest management (IPM) strategy to decrease yield loss caused by pests.

並列關鍵字

tea pest ； lesion identification ； loss function of location regression ； Cascade R-CNN

參考文獻

Bhatt, P. V., Sarangi, S., Pappula, S. (2019, May). Detection of diseases and pests on images captured in uncontrolled conditions from tea plantations. In Autonomous Air and Ground Sensing Systems for Agricultural Optimization and Phenotyping IV (Vol. 11008, p. 1100808). International Society for Optics and Photonics.

Google Scholar

Cai, Z., Vasconcelos, N. (2018). Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 6154-6162).

Google Scholar

Canny, J. (1986). A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6), 679-698.

Google Scholar

Chen, J., Liu, Q., Gao, L. (2019). Visual tea leaf disease recognition using a convolutional neural network model. Symmetry, 11(3), 343.

Google Scholar

DeChant, C., Wiesner-Hanks, T., Chen, S., Stewart, E. L., Yosinski, J., Gore, M. A., ... Lipson, H. (2017). Automated identification of northern leaf blight-infected maize plants from field imagery using deep learning. Phytopathology, 107(11), 1426-1432.

Google Scholar

國際替代計量

區域卷積神經網路及位置回歸方法於茶葉病蟲害辨識之應用

全文下載

主題瀏覽