YOLOv8-MLD: A Lightweight Method for Weed Detection in Cornfields

Le résumé fourni par la source

In the context of Agriculture 4.0, the intelligence of weed control in cornfields still needs more lightweight models to solve the problems of low efficiency and high cost of traditional manual detection and computational complexity of existing deep learning models. In order to realize the lightweight weed detection model, this paper proposes YOLOv8-MLD based on YOLOv8. First, we introduce the C2f-ML module utilizing mixed local channel attention (MLCA) that reducing the network parameters and computational redundancy while maintaining the detection accuracy. Secondly, an ML-BiFPN neck network is constructed by integrating the C2f-ML module into the bidirectional feature pyramid network (BiFPN), augmenting the capability for multi-scale feature fusion. Furthermore, the task dynamic align detection head (TDADH) is designed to fully leverage multi-level features for improved target localization and classification; the minimum point distance-IoU (MPDIoU) loss function is employed to augment the bounding box regression capability. Ultimately, to enhance the model's efficiency, we apply knowledge distillation to facilitate the training of the pruned YOLOv8-MLD. Experiments show that the enhanced YOLOv8-MLD (20% pruning + distillation) achieves a mAP (0.5) of 96.0%, marking a 4.4% improvement over the YOLOv8n baseline, while reducing parameters by 60%, FLOPs by 39.5%, and model size by 53%. The improved model achieves an optimal balance between lightweight and detection accuracy, which helps to reduce deployment costs and makes it easier to implement on UAVs.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.