YOLO-TP: A lightweight model for individual counting of Lasioderma serricorne

Abstract

The quality and safety of tobacco storage and production processes are significantly compromised by the presence of tobacco beetles(Lasioderma serricorne). Currently, the detection of these pests relies on manual counting, a method that is both time-consuming and labor-intensive. Given the limitations of hardware equipment, developing a model that is both easy to deploy and lightweight is especially crucial. To address this need, a new model based on the YOLOv8n architecture, named YOLO-TP, has been specially designed and introduced. The model incorporates the Grouped Shuffle Convolution (GSConv) and an optimized new PC2f structure with Partial Convolution (PConv), aimed at reducing redundant channel computations. Additionally, by employing the Generalized Intersection over Union (GIoU) loss function, it collectively achieves the goals of performance optimization and model lightweighting. YOLO-TP has achieved a high accuracy rate of 99.5% on the tobacco beetle dataset, while simultaneously reducing model parameters and computational requirements by 57.81% and 46.34%, respectively. Compared with existing advanced models, YOLO-TP maintains its lightweight advantage while demonstrating superior performance, offering valuable insights for the development of target detection technology in tobacco beetles and similar fields.