RFDNet-YOLOv13: A lightweight YOLOv13n-based model with Receptive-Field and Dynamic Head for robust tobacco leaf disease detection

• Novel lightweight model: RFDNet-YOLOv13 integrates RFAConv and DyHead modules into YOLOv13 for robust tobacco disease detection. • Enhanced feature extraction: RFAConv improves local receptive field modeling for irregular lesions under complex backgrounds. • Dynamic multi-scale alignment: DyHead adaptively fuses features across scale, spatial, and task dimensions. • Superior accuracy: Achieves 74.7% mAP@50-95 and 81.9% mAP@75, outperforming YOLOv8, YOLOv11, and RT-DETR-L. • Real-time performance: Maintains 154 FPS inference speed, suitable for edge deployment in field environments. • Robust to challenging conditions: Validated under diverse phenological scenarios including rain, fog, and variable lighting. • Synergistic module design: Ablation studies show RFAConv and DyHead work complementarily, yielding gains beyond individual use. Early identification of tobacco diseases is crucial for safeguarding yield, ensuring quality, and minimizing economic losses. However, field-based detection encounters significant challenges due to the small scale, irregular morphology, and fine texture of disease spots, compounded by complex and variable phenological conditions. Traditional manual inspection methods are inefficient, subjective, and inadequate for the demands of modern precision agriculture. To address these challenges, this study proposes a high-precision tobacco leaf disease detection algorithm named RFDNet, based on YOLOv13. First, YOLOv13 was selected as the competitive baseline model through systematic comparative experiments. Subsequently, the RFAConv (Receptive-Field Attention Convolution) module was integrated into the feature fusion neck to enhance the model's capability for feature extraction of multi-scale, irregular lesions and to improve its resistance to background interference by leveraging spatial attention mechanisms. Furthermore, the detection head was reconstructed using DyHead (Dynamic Head), which enhances feature alignment and representation in complex scenes by dynamically integrating scale-aware, spatial-aware, and task-aware attention. Additionally, a 'coarse-to-fine' two-stage training strategy was developed, incorporating a low-learning-rate fine-tuning phase after standard training to overcome convergence bottlenecks in high-precision intervals. Experimental results demonstrate that the proposed method achieves superior performance on the test set, with a final mAP@50-95 of 74.7% and mAP@75 of 81.9%, surpassing mainstream detection models. Ablation studies confirm the effectiveness of the architectural improvements, showing a 2.5 % increase in mAP@50-95 over the baseline. The model effectively addresses the issues of missed and false detection in complex field environments, providing a reliable solution for the intelligent and lightweight deployment of tobacco disease monitoring systems.