Fine-Grained Classification of Wolfberry Pests Based on Generative Self-Supervised Learning

Fine-grained classification of wolfberry pests is crucial for effective pest monitoring and precise control. To address the challenges of fine-grained classification, such as subtle inter-class differences, significant intra-class variations, complex backgrounds, and limited annotated data, this paper proposes PAFT-WPest, a novel method based on generative self-supervised learning. Our approach employs a partial convolutional spatial attention mechanism to focus on insect body regions, combines channel semantic selection with frequency-domain modeling to enhance sensitivity to fine details, and incorporates structural dependency modeling to improve the understanding of global pest semantics. Furthermore, we introduce two fine-grained wolfberry pest datasets, WP45 and WP11, which categorize pests by their growth stages and damage locations, enriching the task's semantic granularity. To bolster domain adaptability, a continuous pre-training strategy is adopted, facilitating effective knowledge transfer from general visual representations to pest-specific features. Experimental results show that PAFT-WPest achieves state-of-the-art accuracies of 97.82% and 94.69% on our WP45 and WP11 datasets, respectively, outperforming existing methods. It also demonstrates strong cross-crop generalization on the public IP102 dataset. The proposed method effectively improves fine-grained pest classification in complex environments, showing promising potential for intelligent agricultural pest monitoring and precision control applications.