Lightweight real-time detection of sweet potato virus disease by a hybrid CNN–Transformer network

Timely and precise detection of sweet potato virus disease (SPVD) is essential for ensuring crop health and improving yield quality in precision agriculture. However, conventional deep learning models often struggle to capture subtle lesion features under complex field conditions. This paper proposes Swin-SPVDNet, a hybrid CNN-Transformer detection framework with task-specific architectural enhancements that combines convolutional neural networks (CNNs) with a modified Swin Transformer to effectively integrate local spatial details and global contextual information. Unlike conventional hybrid models that simply stack convolutional and transformer blocks, the proposed network introduces three key structural designs: (1) A C2PSA-enhanced convolutional backbone for improved local feature sensitivity. (2) A C3STR module that tightly integrates convolutional operations with Swin Transformer blocks for joint local–global feature interaction. (3) A Residual Attention Block (RAB) that refines feature representations while stabilizing gradient propagation. A field-collected SPVD image dataset was constructed using handheld RGB imaging under real agricultural conditions. Extensive experiments and ablation studies demonstrate that Swin-SPVDNet achieves competitive performance compared with representative state-of-the-art detectors while providing a favorable balance between detection accuracy and deployment efficiency, attaining a mean Average Precision ([email protected]) of 86.8%, a precision of 85.1%, and a recall of 75.6%. Compared to Yolov5-v11 models, our approach demonstrates improved robustness in detecting small and irregular lesions, particularly under occlusion and lighting variability. These results confirm the practical value of Swin-SPVDNet for real-time field deployment and support its application in intelligent disease monitoring.