基于像素分类的图像语义分割方法及其应用研究  

Research on image semantic segmentation method based on pixel classification and its application

作　　者：梁智宇 苏彩红[1] 林军帆 LIANG Zhi-yu;SU Cai-hong;LING Jun-fan(School of Mechanical Engineering and Automation,Foshan University,Foshan 528000,China)

机构地区：[1]佛山科学技术学院机电工程与自动化学院,广东佛山528000

出　　处：《佛山科学技术学院学报（自然科学版）》2022年第1期48-53,共6页Journal of Foshan University(Natural Science Edition)

基　　金：广东省普通高校科研项目资助(2018KTSCX237,2019KZDZX1034);佛山科学技术学院学生学术基金项目资助(xsjj202002kjb05)。

摘　　要：为了提高药片外观缺陷检测效率,提出一种使用深度学习进行药片外观缺陷检测的基于像素分类的图像语义分割方法。首先,利用全卷积神经网络对预处理后的样本数据集进行训练,提取药片外观的缺陷特征,然后,采用像素精度和交并比来评估模型分割的精确度,最后,使用训练得到的模型分割出药片的缺陷位置并加以语义的描述。实验结果表明,该方法应用在药片外观缺陷检测中有较高的检测精确度,对其他产品的外观缺陷检测具有很好的参考作用。In order to enhance the efficiency of pill appearance defect detection,in this paper,we propose an image semantic segmentation method based on pixel classification that using deep learning to detect pill appearance defect.Firstly,to obtain the pixel classification model,we train the preprocessed sample data and extract the defect features of pill by applying fully convolutional network.Then the pixel accuracy and intersection ratio are used to evaluate the accuracy of the model segmentation.Finally,the defect location of the tablet is segmented and semantically described by the trained model.The results of the experiments suggest that the proposed method can achieve fast speed and high accuracy in pill appearance defect detection.Moreover,this method of appearance defect detection can provide a useful reference for other products.

关 键 词：深度学习 图像语义分割 像素分类 全卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]