Rice pest Detection based on MobileNetv2-YOLOv4

For the management and conservation of rice, accurate pest identification is crucial. Low pest detection effectiveness and good classification accuracy in rice fields are addressed by the lightweight model MobileNetv2-YOLOv4, which is based on YOLOv4. First, MobileNetv2 is utilized as the backbone network in lieu of CSPDarknet53 in YOLOv4 to minimize the amount of parameters and lighten the model. Second, the model emphasizes the difficulty of recognizing samples during training by balancing the positive and negative samples in the picture using the Focal Loss loss function. The results show that MobileNetv2-YOLOv4 has 90.50% mAP, 51.55 frames per second(FPS), and the model weights of 49.00MB on the 3-type field rice pest dataset. Compared to YOLOv4, the mAP is basically the same, FPS is 206% higher, and model weights are reduced fivefold. The MobileNetv2-YOLOv4 model indicated in this research can be used for field detection of rice pests and diseases, taking into account the accuracy and recognition speed of rice pest detection.