Jiao Lin, Li Gaoqiang, Chen Peng, Wang Rujing, Du Jianming, Liu Haiyun, Dong Shifeng
National Engineering Research Center for Agro-Ecological Big Data Analysis and Application, Information Materials and Intelligent Sensing Laboratory of Anhui Province, School of Internet, Anhui University, Hefei, China.
Institute of Intelligent Machines, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, China.
Front Plant Sci. 2022 Jun 2;13:895944. doi: 10.3389/fpls.2022.895944. eCollection 2022.
An accurate and robust pest detection and recognition scheme is an important step to enable the high quality and yield of agricultural products according to integrated pest management (IPM). Due to pose-variant, serious overlap, dense distribution, and interclass similarity of agricultural pests, the precise detection of multi-classes pest faces great challenges. In this study, an end-to-end pest detection algorithm has been proposed on the basis of deep convolutional neural networks. The detection method adopts a deformable residual network to extract pest features and a global context-aware module for obtaining region-of-interests of agricultural pests. The detection results of the proposed method are compared with the detection results of other state-of-the-art methods, for example, RetinaNet, YOLO, SSD, FPN, and Cascade RCNN modules. The experimental results show that our method can achieve an average accuracy of 77.8% on 21 categories of agricultural pests. The proposed detection algorithm can achieve 20.9 frames per second, which can satisfy real-time pest detection.
根据病虫害综合防治（IPM），准确且强大的害虫检测与识别方案是实现农产品高质量和高产量的重要一步。由于农业害虫姿态多变、严重重叠、分布密集以及类间相似性，多类害虫的精确检测面临巨大挑战。在本研究中，基于深度卷积神经网络提出了一种端到端的害虫检测算法。该检测方法采用可变形残差网络提取害虫特征，并采用全局上下文感知模块来获取农业害虫的感兴趣区域。将所提方法的检测结果与其他先进方法（如RetinaNet、YOLO、SSD、FPN和Cascade RCNN模块）的检测结果进行比较。实验结果表明，我们的方法在21类农业害虫上可实现77.8%的平均准确率。所提检测算法每秒可处理20.9帧，能够满足实时害虫检测的需求。
