智能科学与技术学报 ›› 2019, Vol. 1 ›› Issue (1): 88-95.doi: 10.11959/j.issn.2096-6652.201906

• 学术论文 • 上一篇    下一篇

结合深度学习与生物特征识别在冷链拣选中的算法研究

陈妮亚1(),阮佳阳1,黄金苗1,杨伟2   

  1. 1 ABB中国研究院,北京 100015
    2 ABB机器人应用中心,上海 201319
  • 修回日期:2019-02-21 出版日期:2019-03-20 发布日期:2019-05-28
  • 作者简介:陈妮亚(1987- ),女,安徽合肥人,ABB中国研究院研究员,主要研究方向为机器学习、计算机视觉、监测诊断等。|阮佳阳(1987- ),男,安徽六安人,ABB中国研究院研究员,主要研究方向为机器学习、计算机视觉、监测诊断等。|黄金苗(1986- ),女,重庆人,ABB中国研究院研究员,主要研究方向为机器学习、深度学习、计算机视觉等。|杨伟(1982- ),男,山东诸城人,上海ABB工程有限公司工程师,主要研究方向为机器人应用技术、机器视觉等。

Algorithm design for food-picking combining deep learning and biometrics recognition

Niya CHEN1(),Jiayang RUAN1,Jinmiao HUANG1,Wei YANG2   

  1. 1 ABB Corporate Research Center,Beijing 100015,China
    2 ABB Robotics Application Center,Shanghai 201319,China
  • Revised:2019-02-21 Online:2019-03-20 Published:2019-05-28

摘要:

机器人对食品的抓取、拣选是工业自动化中的一个常见问题,但是食品通常形状不规则、特征多变,导致快速、稳定的视觉分析与定位较为困难。提出了一种结合深度学习与生物特征识别的目标点定位方法,首先使用深度学习模型对目标点进行粗定位,之后在粗定位点的邻域中,利用生物特征进行微调,得到精准的目标点坐标。所设计的方法在常见食品——虾的数据上进行了模型训练与性能验证。首先将虾的图像进行预处理后输入深度学习模型得到粗定位点,之后对虾的位姿进行归一化并提取轮廓线,基于对搜索域内的轮廓拟合与特征点检测以精确定位目标点。实验结果证明了该方法的有效性:在包含 1000 张实测样本的测试集上,整体方案的识别率达到97%,可初步满足实际工业应用的要求。

关键词: 机器视觉, 食品拣选, 深度学习, 特征识别

Abstract:

Irregular-shape food picking and processing is a common problem in industrial automation,which can be difficult for classical image processing techniques,because of the big variations in food shape and characteristics,also the high-performance requirements in both algorithm accuracy and speed.In this paper,a hybrid method based on deep learning and feature recognition was proposed,which first roughly localized the target point based on deep learning model,and then created a search range accordingly.After that,based on biological feature analysis,target point could be accurately localized in the search range.Based on the data of shrimps,a kind of common food,the performance of the proposed method was tested.The shrimp images were pre-processed and used to train the deep learning model for rough localization.Then the shrimp body pose was normalized for edge extraction after proper rotation and projection.The extracted edge curve in search range was analyzed to accurately localize the target joint point.The validation results based on a test set including 1000 samples prove the feasibility of the method – the final detection rate of the proposed hybrid method is 97%.The performance meets industrial requirements on this case.

Key words: computer vision, food picking, deep learning, feature recognition

中图分类号: 

No Suggested Reading articles found!