摘要
与异色背景相比,同色背景下目标的识别特征较少,更易受到遮挡和复杂背景的干扰导致同色背景下对黄瓜果实的识别检测一直是研究领域的重难点之一。针对该问题,提出了一种基于YOLOv11n的目标检测网络YOLO-ACG。引入自适应动态下采样(Adaptive Dynamic Downsample,ADown)模块,融合可变形卷积和通道注意力机制,实现跨尺度特征自适应采样;构建Ghost_HGNetV2网络结构,其中高分辨率组卷积(High-resolution Group Stem,HGStem)将输入图像的通道数压缩,生成固有的特征映射,实现高效特征提取,Ghost_HGBlock模块采用知识蒸馏技术增强特征表达能力;引入上下文与空间特征校准网络结构(Context and Spatial Feature Calibration Network,CSFCN),该结构包含上下文特征校准(Context Feature Calibration,CFC)和空间特征校准(Spatial Feature Calibration,SFC),通过聚合每个像素相关的上下文信息,利用校准空间特征,确保网络正确理解图像的空间布局,进而更加精确地区分具有相似颜色的黄瓜果实和背景。经实验验证,改进后模型精确率提高4.64个百分点,召回率提高5.07个百分点,F1提高4.89个百分点,mAP值提高4.48个百分点。消融、对比实验表明,YOLO-ACG在同色背景下黄瓜果实识别中明显减少了误检和漏检的问题,且具有更高的检测精度。
Compared with a different-color backgrounds,recognizing and detecting cucumber fruits under uniform-color backgrounds remains a key challenge due to limited distinguishing features and increased susceptibility to occlusion and background interference.To address this,we propose YOLO-ACG,a detection network based on YOLOv11n.An Adaptive Dynamic Downsample(ADown)module is introduced,combining deformable convolution and channel attention to achieve adaptive cross-scale feature sampling.A Ghost_HGNetV2 architecture is designed,where the High-resolution Group Stem(HGStem)reduces input channels to extract efficient intrinsic features,and the Ghost_HGBlock applies knowledge distillation to enhance feature representation.A Context and Spatial Feature Calibration Network(CSFCN)network structure is introduced,which includes Context Feature Calibration(CFC)and Spatial Feature Calibration(SFC).The CFC module aggregates context information relevant to each pixel,while the SFC module leverages calibrated spatial features to ensure accurate understanding of spatial layout the image.Together,they enable the network to more precisely distinguish cucumber fruits from backgrounds with similar colors.Experimental results show that the improved model achieves 4.64 percentage points increase in precision,recall by 5.07 percentage points,F1 by 4.89 percentage points,and mAP by 4.48 percentage points.Ablation and comparative experiments confirm that YOLO-ACG significantly reduces false positives and missed detections,offering effective technical support for cucumber fruits recognition in complex,uniform-color environments.
作者
曹丽英
刘洋
王喜林
周恒宇
姜冬辉
CAO Liying;LIU Yang;WANG Xilin;ZHOU Hengyu;JIANG Donghui(College of Information Technology,Jilin Agricultural University,Changchun 130118,China)
出处
《无线电通信技术》
北大核心
2025年第5期1036-1045,共10页
Radio Communications Technology
基金
吉林省科技发展计划项目(20250601061RC)
吉林省教育厅科学研究项目(JJKH20250571KJ)。
关键词
特征提取
YOLO
果实识别
目标检测
feature extraction
YOLO
fruit identification
object detection
