Objective To address the challenge in tea leaf disease detection, where existing deep learning models struggle to balance accuracy and efficiency, particularly unsuitable for deployment on resource-constrained embedded devices.

Method A lightweight and high-accuracy recognition model of MobileNetV2-GCA-LS was proposed based on the MobileNetV2 architecture, incorporating two key innovations. First, a novel ghost coordinate attention (GCA) module was designed, which integrated the positional sensitivity of coordinate attention with the computational efficiency of GhostNet, thereby enhanced the feature representation of critical disease areas. Second, the label smoothing (LS) regularization strategy was employed to optimize the training process and improve the model generalization capability. The model was trained and validated on a publicly available tea leaf disease dataset.

Result The proposed MobileNetV2-GCA-LS model achieved a recognition accuracy of 94.54% and F₁ of 94.29% on the test set, significantly outperforming comparison models including MobileNetV2, MobileNetV3-Small, EfficientNet-B0, ResNet50 and GhostNet. Meanwhile, the model maintained low complexity with 2.608 9×10⁶ parameters and 0.334 7×10¹⁰ floating point operations (FLOPs) of computational cost, demonstrating its feasibility for deployment on resource-constrained devices.

Conclusion The proposed model effectively enhances the performance of tea leaf disease recognition, achieving a well-balanced trade-off between accuracy and efficiency. This provides a practical technical solution for intelligent monitoring and precise control of plant diseases in smart agriculture.