CT-PestNet: A novel crop pest classification model based on cross-scale feature distillation enhancement and vision...

The continuous growth of the global population has led to an increasing demand for food, yet the annual food losses caused by pests are immeasurable. Therefore, accurately identifying crop pests and adopting effective control strategies is of great practical significance. In recent years, intelligent crop pest classification models based on deep learning have made significant progress, which can be primarily divided into two categories: convolutional neural network (CNN)-based and Transformer-based. However, the ability of CNN is limited in modeling the relationship between long-distance regions in images; the capability of Transformer is insufficient in capturing the local discriminative features of crop pests, which leads to unsatisfactory performance of the above two intelligent crop pest classification models. To address the above issues, this paper proposes a new crop pest classification model (CT-PestNet). First, CT-PestNet combines CNN and Transformer to fully leverage the former's strengths in local detail extraction and the latter's capabilities in modeling global dependencies. Second, a cross-scale feature distillation enhancement module (CFDE) is designed to respond to the challenges of diverse pest morphologies (with varying sizes and shapes) and complex backgrounds. The CFDE employs the "distillation-then-enhancement" strategy to enhance the model's perception of small or diverse pests, thereby improving its classification accuracy and robustness in complex backgrounds. Finally, a local token interaction module (LTIR) is constructed to overcome the insufficient modeling of adjacent token features in the Transformer encoder. The LTIR can achieve deep discriminative feature mining by enhancing the feature interaction among adjacent tokens, thereby improving the model's processing ability and recognition performance for complex images. On the IP102 and D0 datasets, CT-PestNet achieves accuracies of 77.32% and 99.73%, respectively; especially on the more challenging IP102 dataset, it outperforms the best CNN-based model by 3.62% and the best Transformer-based model by 1.32%, fully proving its effectiveness.