摘要
针对复杂农业场景下蓝莓果实检测存在的遮挡、光照不均及小目标漏检等问题,传统的目标检测模型存在小目标识别能力不足和多特征融合效果不佳的缺陷,文章提出一种专门用于蓝莓检测任务的BCD-YOLO模型。该模型是基于YOLO系列最新迭代版本YOLOv11的改进版,融入了多项创新设计。该模型在YOLOv11基础上引入BIMAFPN双向特征金字塔网络,通过多层级的特征金字塔以及双向信息传递来提高蓝莓识别精度并设计了CSP-PMSFA模块优化主干网络。该模块采用高效的部分多尺度特征提取,能够从输入中提取多种尺度的特征信息,实现增强的特征融合。此外,该模型设计并加入了基于注意力机制的目标检测头DyHead。实验表明,在自建蓝莓数据集上,改进后的BCD-YOLO模型在mAP50上达到了59%,P(精度)和mAP50-95较原模型YOLOv11分别提升了2.7%和3.1%,同时模型参数量和计算量的增加保持在合理范围内。总体而言,该模型为果园自动化采收和产量预测提供了高鲁棒性的技术解决方案。
Aiming at the problems of occlusion,uneven lighting,and missed detection of small targets in blueberry fruit detection under complex agricultural scenarios,traditional object detection models suffer from insufficient small-target recognition capabilities and poor multi-feature fusion effects.This paper proposes a BCD-YOLO model specifically designed for blueberry detection tasks.Based on the latest iteration of the YOLO series,YOLOv11,this model incorporates several innovative designs.BIMAFPN(Bidirectional Feature Pyramid Network)is introduced on the basis of YOLOv11.Through multi-level feature pyramids and bidirectional information transmission,it improves the recognition accuracy of blueberries.Also,the CSP-PMSFA module is designed to optimize the backbone network.This module uses efficient partial multi-scale feature extraction to extract feature information of multiple scales from the input,achieving enhanced feature fusion.Additionally,an innovative detection head based on attention mechanism,DyHead,is integrated into the detection model.Experiments show that on the self-built blueberry dataset,the improved BCD-YOLO model achieves 59% in mAP50,with precision and mAP50-95 increasing by 2.7% and 3.1% respectively compared to the original YOLOv11 model,while the increase in model parameters and computational complexity remains within a reasonable range.Overall,this model provides a highly robust technical solution for automated orchard harvesting and yield prediction.
作者
刘波
LIU Bo(Anhui Technical College of Mechanical and Electrical Engineering,Wuhu 241000,China)
出处
《无线互联科技》
2025年第16期18-23,共6页
Wireless Internet Science and Technology
基金
安徽省高校自然科学研究项目,项目名称:基于深度学习的语义通信系统在动态数据环境下的应用研究,项目编号:2024AH050212
安徽省高校自然科学研究项目,项目名称:基于多模态情感分析的门控循环分层融合网络构建与应用研究,项目编号:2024AH050216。
关键词
蓝莓检测识别
小目标检测
多尺度特征融合
YOLOv11
智慧农业
blueberry detection and recognition
small-target detection
multi-scale feature fusion
YOLOv11
smart agriculture
