Hierarchical Knowledge-Guided Recognition for Wolfberry Pests

As a medicinal material rich in polysaccharides, Lycium barbarum is susceptible to pest infestation, and early and accurate pest identification is crucial for prevention and control. However, Lycium barbarum pest identification belongs to a typical fine-grained classification task, and practical applications often face the challenge of sample scarcity. Existing methods perform insufficiently in the utilization of long text and the modeling of confusion under few-shot conditions. To address these issues, this paper constructs a Hierarchical Knowledge-Guided Framework (HKGF) based on Long-CLIP. This framework fully utilizes the alignment between long textual descriptions of pests and multi-scale visual features. Furthermore, a knowledge-guided de-confusion module is introduced to explicitly model the confusion relationships between categories using textual and visual similarity graphs. Under the few-shot setting of the Ningxia Lycium barbarum pest dataset, the Top-1 accuracy reached 85.32% and the Top-5 accuracy reached 97.92%. Compared with standard CLIP, it improved by 42.58%; compared with Long-CLIP, it improved by 31.46%; and compared with LDC+Long-CLIP, it improved by 1.72%, surpassing existing baseline methods. HKGF can fully mine the complementary information in pest images and descriptions, effectively solving the fine-grained confusion problem under data-constrained conditions, thus providing an effective solution for pest identification and intelligent prevention and control.