Development of an automatic pest monitoring system using a deep learning model of DPeNet

Abstract

Pest monitoring is the prerequisite for precise decision-making supporting integrated pest management (IPM) in smart agriculture. In order to satisfy the timeliness and accuracy in practical application, an automatic detection system is proposed, consisting of a camera, a micro-computer, a deep learning model, and a server. For compatibility of multi-scale objects, a convolutional neural network with a brand new self-designed backbone DPeNet is presented. The DPeNet splits residual blocks into aggregated residual blocks to learn more features with fewer parameters. In addition, the DPeNet applies a deformable convolutional network to weaken the influence of shape changes of pests. The improved pest detection network with DPeNet is adequate for simultaneously detecting insects of different sizes. Results indicate that DPeNet achieved Average Precision (AP) of 0.932, outperforming benchmark methods including Faster R-CNN, SSD, and Yolov3. Acceptable performances are obtained in practical experiment tests, which satisfy the actual work demands.