PKD-YOLOv8: A Collaborative Pruning and Knowledge Distillation Framework for Lightweight Rapeseed Pest Detection

As an important oil and vegetable crop, rapeseed is widely planted and has important economic value worldwide. Rapeseed is often threatened by various pests during its growth. In order to effectively deal with rapeseed pests, this paper proposes a lightweight method based on collaborative compression learning. This method uses YOLOv8s as the basic model, combines model structure analysis and pruning sensitivity evaluation, and implements structured pruning to compress the model size. The Logit distillation method is integrated with the improved generative distillation method MGD, and the LMGD distillation strategy is proposed to enhance the student model's ability to fit the teacher model's feature expression. In order to verify the effectiveness of the proposed method, we built a rapeseed pest dataset (ACEFP) and conducted experiments. The improved model achieved 96.7% mAP@0.5, 93.2% accuracy, and 92.7% recall, while the parameter size was compressed from 11.2 MB to 4.4 MB, and the FLOPs were reduced from 28.3 G to 10.01 G, which were reduced by about 60.7% and 64.6%, respectively, and the accuracy was only reduced by 0.1%. The model achieved a measured frame rate of 11.76 FPS on the Jetson Nano edge device, demonstrating excellent real-time inference performance.