Swin Attention Augmented Residual Network: a fine-grained pest image recognition method

Pest infestation is a major cause of crop losses and a significant factor contributing to agricultural economic damage. Accurate identification of pests is therefore critical to ensuring crop safety. However, existing pest recognition methods often struggle to distinguish fine-grained visual differences between pest species and are susceptible to background interference from crops and environments. To address these challenges, we propose an improved pest identification method based on the Swin Transformer architecture, named Swin-AARNet (Attention Augmented Residual Network). Our method achieves efficient and accurate pest recognition. On the one hand, Swin-AARNet enhances local key features and establishes a feature complementation mechanism, thereby improving the extraction capability of local features. On the other hand, it integrates multi-scale information to effectively alleviate the problem of fine-grained feature ambiguity or loss. Furthermore, Swin-AARNet attained a classification accuracy of 78.77% on IP102, the largest publicly available pest dataset to date. To further validate its effectiveness and generalization ability, we conducted additional training and evaluation on the citrus benchmark dataset CPB and Li, achieving impressive accuracies of 82.17% and 99.48%, respectively. SwinAARNet demonstrates strong capability in distinguishing pests with highly similar appearances while remaining robust against complex and variable backgrounds. This makes it a promising tool for enhancing agricultural safety management, including crop environment monitoring and early invasion warning. Compared with other state-of-the-art models, our proposed method exhibits superior performance in pest image classification tasks, highlighting its potential for real-world agricultural applications.