基于改进YOLOv10n的马铃薯形状识别方法OA
A potato shape recognition method based on an improved YOLOv10n
为提高马铃薯采后分选过程中畸形薯识别准确率,解决人工分选效率低、主观性强以及传统机械分级设备难以识别复杂畸形形态等问题,本研究提出一种融合全局通道空间注意力(global channel-spatial attention,GCSA)、多尺度空洞特征聚合模块(multi-scale dilated feature aggregation,MDFA)和中值增强空间通道注意力(median-enhanced spatial and channel attention,MECS)模块的改进YOLOv10n马铃薯形状识别方法.以正常马铃薯和典型畸形马铃薯为研究对象,利用机器视觉图像采集平台构建 normal 和 deformity 两类数据集.首先选取 YOLOv8n、YOLOv10n、YOLO11n 和YOLO12n进行基础模型对比试验,筛选综合检测性能较优的YOLOv10n作为基础网络;随后在YOLOv10n中引入全局通道空间注意力模块、多尺度空洞特征聚合模块和中值增强空间通道注意力模块,分别增强模型对关键区域、多尺度形态特征和复杂边界特征的提取能力,并通过消融试验验证改进效果.数据集标注结果显示,畸形马铃薯样本数量为2 175个,正常马铃薯样本数量为738个.基础模型对比结果表明,YOLOv10n的精确率、召回率和平均精度均值(mean average precision,mAP@0.5)最大值分别为91.500%、90.990%和90.980%,综合性能优于其他对比模型.消融试验结果表明,YOLOv10n+GCSA+MDFA+MECS 模型取得较优识别效果,其精确率、召回率、mAP@0.5 分别达到 97.823%、96.835%和96.890%,mAP@0.5较原始YOLOv10n提高5.910个百分点.改进模型能够有效增强马铃薯异常轮廓、局部凸起和多尺度形态特征的表达能力,可为马铃薯智能分选中畸形薯自动识别提供算法基础.
To improve the recognition accuracy for normal and deformed potatoes and enhance the model's ability to extract irregular contours,local protrusions,and complex boundary features,this study proposes an improved YOLOv10n-based potato shape recognition method integrating GCSA,MDFA,and MECS modules.Normal potatoes and typical deformed potatoes were used as the research objects,and a machine vision image acquisition platform was used to construct two datasets:normal and deformity.On this basis,YOLOv8n,YOLOv10n,YOLO11n,and YOLO12n were selected for comparative experiments,and YOLOv10n was chosen as the baseline model.Subsequently,a global channel-spatial attention module,a multi-scale dilated feature aggregation module,and a median-enhanced spatial and channel attention module were incorporated into the YOLOv10n network to improve the model in terms of key region attention,multi-scale feature fusion,and complex boundary feature enhancement,respectively.Ablation experiments were conducted to verify the model performance.The dataset annotation results showed that there were 2 175 deformed potato samples and 738 normal potato samples.The comparison results of the baseline models showed that the maximum precision,recall,and mAP@0.5 of YOLOv10n reached 91.500%,90.990%,and 90.980%,respectively,outperforming the other comparison models.The ablation experiment results showed that the YOLOv10n+GCSA+MDFA+MECS model achieved the best recognition performance,with precision,recall,and mAP@0.5 reaching 97.823%,96.835%,and 96.890%,respectively.Compared with the original YOLOv10n,the mAP@0.5 increased by 5.910 percentage points.The improved YOLOv10n model can effectively enhance the representation ability of abnormal contours,local protrusions,and multi-scale morphological features of potatoes,improve the recognition performance of normal and deformed potatoes,and provide an algorithmic basis for intelligent potato sorting and automatic recognition of deformed potatoes.
徐秋颖;汪迁;范文浩;王文渊;邓伟刚
内蒙古农业大学机电工程学院,内蒙古 呼和浩特,010018内蒙古农业大学机电工程学院,内蒙古 呼和浩特,010018内蒙古农业大学机电工程学院,内蒙古 呼和浩特,010018内蒙古农业大学机电工程学院,内蒙古 呼和浩特,010018内蒙古农业大学机电工程学院,内蒙古 呼和浩特,010018
农业科技
马铃薯形状识别机器视觉YOLOv10n注意力机制智能分选
potatoshape recognitionmachine visionYOLOv10nattention mechanismintelligent sorting
《智能化农业装备学报(中英文)》 2026 (3)
30-37,8
内蒙古农业大学高层次/优秀博士引进人才科研启动项目(NDYB2021-12)内蒙古自治区自然科学基金(2026MS0805)内蒙古自治区一流学科科研专项项目(YLXKZX-NND-047)国家自然科学基金(52565026) Scientific Research Start-up Foundation for High-level/Outstanding Doctoral Talents of Inner Mongolia Agricultural University(NDYB2021-12)Natural Science Foundation of Inner Mongolia Autonomous Region of China(2026MS0805)Inner Mongolia Autonomous Region First-Class Discipline Research Special Project(YLXKZX-NND-047)National Natural Science Foundation of China(52565026)
评论