Star-YOLO: A lightweight and efficient model for weed detection in cotton fields using advanced YOLOv8 improvements

Abstract

Effective weed management in cotton fields is crucial for preventing crop loss and maintaining agricultural productivity. However, the complexity and high computational demands of deep-learning models pose challenges when deployed in resource-constrained devices. Hence, this study proposes a lightweight deep-learning model based on an improved YOLOv8 architecture. First, the backbone and C2f modules are restructured using Star Blocks, along with a designed lightweight detection head, i.e., the lightweight shared convolutional separable BN detection head, thus effectively reducing the model’s parameters and computational overhead. To better capture the global weed information, the LSK attention mechanism expands the receptive field, thus enhancing the detection performance of the model. Additionally, a dynamic upsampling technique, DySample, is employed to replace conventional upsampling operators, thereby further improving the detection speed. Compared with YOLOv8, the proposed model reduces the parameters, computation, and model size by 50.0%, 39.0%, and 47.0%, respectively, while achieving mAP@50 and mAP@50–95 scores of 98.0% and 95.4%, respectively. Furthermore, the model optimally balances accuracy, lightweight design, and detection speed compared with mainstream lightweight backbone networks and architectures, thus demonstrating its superior performance on public datasets CottonWeedDet12 and CottonWeedDet3. By integrating TensorRT technology, the model’s detection speed increases by nine times, thus providing significant advancements toward the development of an efficient weed-detection system for real-world agricultural applications. As this model can be integrated into automated weeding equipment, fully automated weed detection and weeding operations are realizable, thereby enhancing the efficiency and precision of agricultural tasks.