Rice Field Pest Detector Based on Deep Learning and Embedded System

To improve the efficiency of rice field pest identification, reduce labor costs, and solve the problem of one-sided judgment in manual detection, a pest detection method based on deep learning and an embedded system was proposed and a portable instrument was designed. The detection model was first obtained by training 1500 rice field image datasets of pests through a YOLOv2-MobileNet network, which was then deployed into a device, and real-time detection of common rice paddy pests was realized. To reduce the impact of model deployment on the device performance, the detection accuracy and image capture frame rate were improved by adjusting the width of the MobileNet network. The results showed that the average detection accuracy of the model trained by the backbone network MobileNet-0.75 was 89.4%, 80.7%, 90.0%, 81.9%, and 83.6% for the locust, rice leaf roller, rice stem borer, rice green mirid nymph, and rice green adult, respectively, and the average image acquisition frame rate reached 35 frames per second. Real-time detection requirements were met. The design realizes real-time detection of five common rice field pests using embedded equipment and provides a reference value for the application of intelligent equipment for the detection of agricultural pests.