TD-YOLO: Tomato Leaf Disease Detection Based on Hypergraph Multi-Scale Aggregation

To address the challenges of tomato leaf diseases, including significant variations in disease size, discrete spatial distribution in complex field environments, and the difficulty in distinguishing early symptoms of certain diseases, which lead to low detection accuracy and a large number of model parameters, this paper proposes a tomato leaf disease detection model, TD-YOLO, based on multi-scale hypergraph aggregation. During the feature extraction stage, the RetC3k2 is employed to enhance the shallow network's ability to detect disease regions. The C2TSSA attention mechanism is introduced to filter out key diseases and suppress background interference. At the same time, the GDown adaptive downsampling module is designed to reduce the model size while preserving feature integrity. In the neck network section, a multi-scale hypergraph aggregation structure is constructed by integrating a multi-scale semantic aggregation layer, the HyperConv, and a MANet network. By connecting disease features at different spatial locations through hyperedges, the network models high-order associative relationships and achieves a refined representation of multi-scale features. Experimental results show that on the Tomato-Village-Diseases dataset, TD-YOLO achieves an mAP50 of 92.5%, a 7.1% improvement over the baseline YOLOv11, with only 2.21M parameters and a computational cost of 6.8 GFLOPs. On the Tomato-Leaf-2-v1 dataset, the mAP50 reached 81.0%, outperforming all comparison models and demonstrating the model's strong adaptability. This study provides an efficient and feasible technical solution for the accurate detection of tomato leaf diseases.