Abstract: 【Objective】 To explore the application effectiveness of a lightweight deep neural network based on knowledge distillation for potato disease recognition,providing a reference for disease prevention and control in smart agriculture.【Methods】 A potato disease dataset containing 12 000 images across three categories（healthy,early blight,and late blight） was constructed using the public PlantVillage dataset and self-collected data. This dataset was divided into training,validation,and test sets in a ratio of 7∶1∶2. First,EfficientNet-B3 was trained as the teacher network. Subsequently,a dual-path knowledge transfer framework was constructed,with MobileNetV3-Small serving as the student network. Key discriminative features were extracted through a feature selection mechanism,and an adaptive knowledge distillation strategy was employed for simultaneous knowledge transfer at the feature and prediction layers. The loss weights were dynamically adjusted based on the training state. The student network was trained under the guidance of the teacher network to obtain an improved model that achieves both lightweight architecture and high recognition performance,and was compared with the un-distilled MobileNetV3-Small baseline model.【Results】 The improved model achieved an accuracy of 97.80% for potato disease detection on the test set,representing a 3.20 percentage point improvement over the baseline model and closely approaching the teacher network′s performance level of 98.50%. The model′s parameter count was only 1.60 M and its floating-point operations（FLOPs）were 62.03 M,which were 85.05% and 93.75% less than the teacher network,respectively. Grad-CAM visualization analysis confirmed that the model precisely locates key diagnostic features,such as the concentric rings of potato early blight and the water-soaked lesions of potato late blight,with the heatmap highly matching with the diseased regions.【Conclusion】 The lightweight potato disease recognition model based on adaptive knowledge distillation achieves both high accuracy and low computational complexity. It can effectively recognize the features of potato early blight and late blight,making it suitable for resource-constrained field environments.

Key words: Lightweight deep neural network, Potato disease recognition, Knowledge distillation, Deep learning, Smart agriculture