Lightweight Real-Time Detection Transformer for Tomato Leaf Disease Recognition in Complex Agricultural Scenarios

Accurate detection of tomato leaf diseases is essential for sustainable tomato production. To overcome the limitations of existing detection models, such as large parameter sizes, insufficient accuracy, weak robustness, and poor small-target detection performance, this study proposes a lightweight tomato leaf disease detection algorithm based on an improved Real-Time Detection Transformer (RT-DETR), named Tomato Leaf Diseases RT-DETR (TLD-RTDETR). Specifically, a partial convolution-based PConvBlock is introduced into the backbone to enhance feature extraction while reducing model complexity. In addition, a Coordinate Attention-based hierarchical feature pyramid module (CA_HSFPN) is designed to suppress background interference and strengthen small-target feature representation. Furthermore, a learnable positional encoding strategy is integrated into the feature encoding stage to improve the extraction of critical disease features in complex environments. Experimental results show that TLD-RTDETR achieves an mAP of 92.9%, precision of 94.2%, and recall of 87.5% on the tomato leaf disease dataset, outperforming the RT-DETR-R18 baseline by 2.4%, 0.5%, and 3.3%, respectively. Meanwhile, the model size, parameter count, and computational cost are reduced by 38.1%, 38.1%, and 31.2%. Compared with mainstream methods, the proposed model achieves better detection performance with a more lightweight architecture. Additional visualization, anti-interference, and generalization experiments further verify its robustness and cross-scene adaptability, demonstrating its potential for practical deployment in tomato leaf disease detection.