A Lightweight Crop Pest Detection Algorithm Based on Improved Yolov5s

The real-time target detection of crop pests can help detect and control in time. In this study, we built a lightweight agricultural pest identification method based on modified Yolov5s reconstructed the original backbone network tandem with MobileNetV3 to considerably reduce number parameters model. At same time, ECA attention mechanism was introduced into shallow meet aim effectively enhancing network’s performance by introducing limited parameters. A weighted bidirectional feature pyramid (BiFPN) utilized replace path aggregation (PAnet) neck boost extraction tiny targets. SIoU loss function CIoU increase convergence speed accuracy model prediction frame. updated designated ECMB-Yolov5. conducted experiments eight types common dataset photos, comparative were using methods. final implemented an embedded device, Jetson Nano, for detection, which gave reference further application UAV or unmanned cart systems. experimental results indicated that ECMB-Yolov5 decreased 80.3% mAP 0.8% compared deployed devices reached 15.2 FPS, 5.7 FPS higher than improved 7.1%, 7.3%, 9.9%, 8.4% Faster R-CNN, Yolov3, Yolov4, Yolov4-tiny models, respectively. It verified through study had high while significantly reducing accomplishing detection.