Real-Time Agricultural Pest Detection Enhanced via Hybrid Distillation Framework

We propose an improved version of the YOLOv8 model that enhances real-time agricultural pest detection performance by integrating feature distillation and relation distillation. Feature distillation focuses on transferring intermediate feature maps from the teacher model to the student model, enriching the student's intermediate representation capabilities. Additionally, relation distillation maintains the relative relationships between feature channels, further boosting the model's generalization ability and detection accuracy. In this process, the teacher model constructs a relationship graph between feature channels and conveys this structural information to the student model. This approach not only emphasizes the semantic information of the intermediate features but also incorporates the geometric relationships between channel-level features, enabling the student model to better understand and preserve the inherent structure of the data. Experimental results on the large-scale agricultural pest dataset IP102 demonstrate that this model significantly improves detection accuracy and robustness compared to the traditional YOLOv8 model. The strategy of combining feature and relation distillation showcases its practicality and potential for wide application in agricultural pest detection tasks.