Insect Pest Detection and Identification Method Based on Deep Learning for Realizing a Pest Control System

• Published in: 2020 59th Annual Conference of the Society of Instrument and Control Engineers of Japan (SICE) pp. 709 - 714
• Main Authors: Kuzuhara, Hiroaki, Takimoto, Hironori, Sato, Yasuhiro, Kanagawa, Akihiro
• Format: Conference Proceeding
• Language: English; Japanese
• Published: The Society of Instrument and Control Engineers - SICE 23.09.2020
• Subjects: Agriculture; Deep learning; image processing; insect pest detection; Insects; Object detection; Pest control; Process control; Proposals
• DOI: 10.23919/SICE48898.2020.9240458

Detection of insect pests in the agricultural field, which is useful in achieving smart agriculture, has attracted considerable attention. In particular, automatically monitoring the number of crop insect pests has evolved into key means of managing and optimizing agricultural resources. However, despite the fact that conventional convolutional neural network (CNN)-based approaches have yielded sufficient results for general object detection, few methods have been developed to detect and recognize small objects such as insect pests. In addition, no large dataset exists for pest detection even though CNN-based methods require a large dataset to optimize many parameters. In this study, we propose two-stage detection and identification methods for small insect pests based on CNN. We also present a region proposal network for insect pest detection using YOLOv3 and propose a re-identification method using the Xception model. To train these models, we propose a data augmentation method using image processing.