基于YOLO-SVP的小尺寸蔬菜害虫检测模型研究OA

Effective pest monitoring is crucial for high-quality vegetable cultivation.While deep learning-based pest detection methods excelling at detecting large-and medium-sized pests,they face challenges with small-sized pests.To address the problem,a you only look once(YOLO)-based small-sized vegetable pest detection method was presented,named YOLO-SVP.To emphasize crucial small-sized pest features and improve feature fusion,a dynamic weighting attention(DWA)mechanism was constructed and integrated into the C3k2 block of YOLO 11,yielding a new block denoted C3k2-DWA.Additionally,to preserve critical spatial information during downsampling and reduce the loss of small pest features,a space-to-depth downsampling(SPD-Down)block was proposed.Besides,to alleviate the severe weakness of bounding box regression in the case of small pests,the normalized Wasserstein distance(NWD)loss function was introduced.Experimental simulation on a self-built vegetable pest dataset demonstrated the effectiveness of the proposed YOLO-SVP,which achieved 85.7%F1 score,89.3%mAP50,and 54.9%mAP50:95,outperforming YOLO 11 by 4.5,3.8,and 4.3 percentages points,respectively.For the Frankliniella occidentalis(small-sized pest),the detection performance improved the F1 score,mAP50,and mAP50:95 by 6.3,8.5,and 5.0 percentages points,respectively.This research provided a paradigm for adapting deep learning architectures to challenge small-sized object detection tasks in precision agriculture,which would provide important support for the effective monitoring of vegetable pests.