TRP-Net: A Text-Guided Dual-Modal Network for Potato Leaf Lesion Detection Under Complex Imaging Conditions.

Potato leaf diseases seriously affect potato yield and quality. Accurate recognition and localization of visible lesions associated with potato early blight and late blight can support preliminary disease screening and targeted field inspection. Object detection can identify disease categories and locate lesion regions, providing reliable information for disease management. However, potato leaf diseases usually occur in complex field environments, where illumination changes, shadows, color shifts, leaf texture interference, and imaging blur may weaken lesion feature representation. In addition, early blight and late blight lesions often have small scales, weak boundaries, irregular shapes, and scattered distributions, making detection prone to background misclassification, missed small lesions, and localization deviation. To address these problems, this study proposes TRP-Net, a text-guided dual-modal detection network, for the robust recognition and localization of visible lesions associated with potato early blight and late blight under complex imaging conditions. First, a Text-Anchored Counterfactual Intervention Module (TACIM) is proposed to construct environmentally perturbed images and constrain feature consistency, improving the robustness of disease representations. Second, a Residual Cross-Modal Attention module (RCMA) is introduced to guide visual features with textual semantics and enhance lesion boundaries and fine-grained textures. Third, a P4 Gaussian Position Distribution Auxiliary Loss (PGDAL) is designed to provide position distribution supervision during early training, improving small-lesion localization. Experimental results on the independent test set show that TRP-Net achieves 78.3% Precision, 73.7% Recall, 75.9% F1 score, 80.5% mAP50, and 34.6% mAP50-95. The detector-side network contains 2.79 M parameters and requires 5.2 GFLOPs. The performance gap between mAP50 and mAP50-95 indicates that precise lesion localization under stricter IoU criteria remains challenging.