ToT-Net: A generalized and real-time crop disease detection framework via task-level meta-learning and lightweight multi-scale transformer

Accurate and efficient detection of crop diseases plays a crucial role in advancing smart agriculture. However, existing object detection models often struggle to generalize under complex environmental conditions and face deployment challenges in practical field applications. To overcome these limitations, this study proposes a Task-oriented Transformer Network (ToT-Net), a lightweight and task-adaptive object detection framework based on an improved RT-DETR architecture. ToT-Net incorporates a multi-dimensional agricultural disease image dataset annotated with region, severity, and growth stage information, establishing a realistic and challenging benchmark for disease detection. The network architecture adopts a Window-Decomposed Transformer Encoder (WDFormer) as the backbone, integrating wavelet convolution and deformable attention to enhance spatial-context representation. A lightweight Focal Feature Modulation and Aggregation Neck (FocalLite-Neck) module is also introduced to strengthen multi-scale feature fusion. A task-level meta-learning strategy is further employed to improve generalization across unseen scenarios. Experimental evaluations show that ToT-Net achieves a mean Average Precision of 94.6 % and an F1-score of 0.9053 on the proposed dataset, with a peak inference speed of 102.3 FPS, surpassing mainstream models such as YOLOv11 and RT-DETR. The model also demonstrates robust performance in adverse conditions, including intense illumination, blur, haze, and rain. ToT-Net generalizes well across public datasets, achieving a 6.1 % average improvement in AP50:95 over baseline methods. These results highlight ToT-Net’s potential as a practical and effective real-time crop disease monitoring solution in intelligent agriculture.