Abstract: By incorporating the coordinate attention mechanism into the CenterNet model, the CAM-CenterNet model focused more on channels and positions that had good representation ability for Lasioderma serricorne （hereinafter referred to as tobacco worms）, reducing the interference of impurities such as cut tobacco and tobacco dust. This study compared the tobacco worms detection performance of CAM-CenterNet model, CenterNet model, YOLOv3 model, and Faster R-CNN model using precision, recall, mAP, frames per second （FPS）, and model parameter size as evaluation metrics. The results indicated that the YOLOv3 model performed the best in terms of recall and average accuracy, while the CAM-CenterNet model lagged slightly behind the YOLOv3 model but outperformed other models;in terms of frame rate, the CAM-CenterNet model detected tobacco worms images faster than the YOLOv3 model, with fewer model parameters and lower requirements for device configuration. The CAM-CenterNet model detected a higher number of tobacco worms than the Faster R-CNN model and YOLOv3 model when detecting smaller individuals. The CAM-CenterNet model not only focused more on the target features of tobacco worms, but also effectively suppressed the interference caused by impurities such as cut tobacco leaves and tobacco dust, achieving effective detection of tobacco worms. The CAM-CenterNet model could meet the requirements of cigarette factories for the speed and accuracy of tobacco pest detection, and could provide technical support for tobacco pest control in cigarette factories.

Key words: coordinate attention mechanism, CenterNet model, CAM-CenterNet model, Lasioderma serricorne detection