摘要
针对目前复杂果园环境下苹果目标检测算法存在模型参数量大、计算复杂度高,难以在计算资源匮乏的设备上应用的问题,提出了一种改进YOLOv8的果园复杂环境下轻量化苹果目标检测算法YOLOv8n-Apple。引入骨干网络VanillaNet,减少模型参数量,降低模型复杂度;将原始模型C2f模块替换为C2fGhost模块,通过较少的卷积运算来获得相似特征图进一步减少模型参数;使用轻量级上采样算子CARAFE,避免传统上采样算子语义缺失和感受野过小的问题;由于传统损失函数不能完全捕捉到目标之间的相对位置和大小差异,采用WIoU边界框作为回归损失函数。收集包含远景顺光、远景背光、近景顺光、近景背光等成熟苹果照片共计3120张,从不同角度和背景进行采集,并改进数据增强,避免数据集单个不确定性;本文提出果园环境下改进后的苹果检测模型平均检测精度分别比SSD、Faster R-CNN、YOLOV5、YOLOV7、YOLOV8高7.5个百分点、4.8个百分点、2.2个百分点、3.8个百分点和3.4个百分点,达到90%,检测速度达到286帧,模型大小1.8 MB,比原始模型提高了41帧,模型大小仅有其60.0%。
Addressing the issues of large model parameters and high computational complexity in apple target detection algorithms for complex orchard environments,which hinder application on devices with limited computational resources,an improved and lightweight apple target detection algorithm named YOLOv8n-Apple based on YOLOv8 was proposed.The backbone network,yaniaNet,was introduced to reduce model parameters and complexity.The original C2f module in the model was replaced with the C2fGhost module,which further decreased model parameters by obtaining similar feature maps through fewer convolutional operations.The lightweight upsampling operator CARAFE was utilized to address the issues of semantic loss and excessively small receptive fields associated with traditional upsampling operators.Given that traditional loss functions cannot fully capture the relative position and size differences between targets,the WIoU bounding box was adopted as the regression loss function.A dataset comprising 3120 images of mature apples in various scenarios,including distant and close views under front-light and backlight conditions,was collected from diverse angles and backgrounds,to mitigate potential dataset uncertainties.The improved apple detection model for orchard environments demonstrated an average detection accuracy of 90%,which was 7.5,4.8,2.2,3.8,and 3.4 percentage points higher than SSD,Faster R-CNN,YOLOv5,YOLOv7,and YOLOv8,respectively.The detection speed reached 286 frames per second,and the model size was reduced to 1.8 MB,representing an improvement of 41 frames per second compared to the original model,while occupying only 60.0%of size.
作者
周晖
杨洁
赵祥飞
ZHOU Hui;YANG Jie;ZHAO Xiang-fei(College of Mechanical and Transportation,Southwest Forestry University,Kunming 650224,China)
出处
《科学技术与工程》
北大核心
2025年第6期2274-2283,共10页
Science Technology and Engineering
基金
云南省教育厅科学研究基金(2023J0711/0111723084)
云南省农业联合专项(202301BD070001-001)
云南省专业学位研究生教学案例库建设项目(503210305)
中国学位与研究生教育学会农林学科工作委员会项目(2021-NLZX-YB14/503210401)。
作者简介
第一作者:周晖(1996-),男,汉族,湖北黄梅人,硕士研究生。研究方向:图像处理、计算机视觉。E-mail:718196955@qq.com;通信作者:杨洁(1973-),女,汉族,河南光山人,博士,副教授。研究方向:机电一体化、机器视觉。E-mail:351725623@qq.com。
