HDR-YOLO: A Lightweight Tomato Leaf Disease Detection Model Based on High-Level Semantic Guidance and Dynamic Upsampling

Tomato leaf diseases seriously affect tomato production, and early image-based screening is important for assisted diagnosis under greenhouse or semi-controlled cultivation conditions. Traditional manual inspection methods are often inefficient and highly subjective. To address these limitations, this study proposes a lightweight tomato leaf disease detection model named HDR-YOLO, which integrates high-level semantic guidance and dynamic upsampling. First, a High-level Semantic Feature Pyramid Network (HSFPN) is introduced to reconstruct the feature fusion neck. This modification significantly reduces the parameter count while enhancing the fusion of semantic features. Second, a dynamic upsampling operator (DySample) replaces traditional upsampling techniques to minimize information loss during feature map expansion. Finally, Receptive Field Coordinate Attention Convolution (RFCAConv) is incorporated into the critical feature fusion path. This enhances spatial coordinate perception, thereby improving the recognition accuracy for diseases with complex textures, such as early blight. Experimental results demonstrate that HDR-YOLO achieves a mean average precision at an Intersection over Union of 0.5 (mAP(0.5%)) of 79.5% on the test dataset. Compared to the baseline YOLOv12n model, the proposed method increases mAP(0.5%) by 2.2% while reducing the number of parameters by approximately 29.5% and computational complexity by approximately 14.3%. HDR-YOLO achieves efficient and accurate detection, providing a lightweight image-based framework for tomato leaf disease screening and assisted diagnosis under greenhouse or semi-controlled cultivation conditions.