Abstract
INTRODUCTION: Timely and accurate recognition of tomato diseases is crucial for improving tomato yield. While large deep learning models can achieve high-precision disease recognition, these models often have a large number of parameters, making them difficult to deploy on edge devices. To address this issue, this study proposes an ensemble self-distillation method and applies it to the lightweight model ShuffleNetV2. METHODS: Specifically, based on the architecture of ShuffleNetV2, multiple shallow models at different depths are constructed to establish a distillation framework. Based on the fused feature map that integrates the intermediate feature maps of ShuffleNetV2 and shallow models, a depthwise separable convolution layer is introduced to further extract more effective feature information. This method ensures that the intermediate features from each model are fully preserved to the ensemble model, thereby improving the overall performance of the ensemble model. The ensemble model, acting as the teacher, dynamically transfers knowledge to ShuffleNetV2 and the shallow models during training, significantly enhancing the performance of ShuffleNetV2 without changing the original structure. RESULTS: Experimental results show that the optimized ShuffleNetV2 achieves an accuracy of 95.08%, precision of 94.58%, recall of 94.55%, and an F1 score of 94.54% on the test set, surpassing large models such as VGG16 and ResNet18. Among lightweight models, it has the smallest parameter count and the highest recognition accuracy. DISCUSSION: The results demonstrate that the optimized ShuffleNetV2 is more suitable for deployment on edge devices for real-time tomato disease detection. Additionally, multiple shallow models achieve varying degrees of compression for ShuffleNetV2, providing flexibility for model deployment.