基于Hu不变矩和径向基神经网络的太阳镜镜片瑕疵图像分类系统  

Image-classification system of sunglass-lens-defects based on Hu-RBF

作　　者：王昕 杨钰萍 何嘉玮 廖志鹏 黄宇宸 范贤光 许英杰 WANG Xin;YANG Yuping;HE Jiawei;LIAO Zhipeng;HUANG Yuchen;FAN Xianguang;XU Yingjie(School of Aerospace Engineering,Xiamen University,Xiamen 361102,China)

机构地区：[1]厦门大学航空航天学院,福建厦门361102

出　　处：《厦门大学学报（自然科学版）》2023年第4期629-637,共9页Journal of Xiamen University：Natural Science

摘　　要：现有的镜片瑕疵分类主要针对光学玻璃镜片和树脂镜片,与太阳镜镜片的生产原料和工艺存在差异,导致两者的瑕疵类型也有所不同,因此现有的镜片检测机不能直接应用于太阳镜镜片的检测.为了满足生产需求,实现工业自动化检测,本文利用基于机器视觉检测技术的CMOS传感器工业相机、双远心镜头、LED灯搭建了镜片图像采集系统,并融合图像处理算法,将七个Hu矩不变量作为互补特征,采用径向基神经网络模型开发了太阳镜镜片瑕疵图像分类系统.实验表明,使用Hu不变矩特征提取算法的分类方法可有效提高分类准确率,该方法的准确率为94.56%.和BP等其他神经网络结构相比,准确率也更高.因此,该系统被证明是可行和有效的.Presently,lens defects are classified primarily by means of detecting optical glass lenses and resin lenses,of which production materials and processes differ from those of sunglasses lenses,resulting in difference of defect types.Consequently,existing lens detection machines cannot be applied to the detection of sunglasses lenses.In this study,to meet the production demand and realize industrial automatic detection,we use CMOS sensor industrial camera,dual telecentric lens and LED lamp based on machine vision detection technology to build a lens image acquisition system and fusion image processing algorithm.Then,seven Hu moment invariants are taken as complementary features,and the RBF neural network is used as the model to develop an image-classification system of sunglass-lens-defect.Experiments show that the classification method jointly with Hu moment invariant feature extraction algorithm can effectively increase the classification accuracy to 94.56%.In comparison with other neural network structures such as BP,the accuracy rate is higher than other neural networks do.Therefore,the proposed system is proved to be feasible and effective.

关 键 词：太阳镜镜片 瑕疵分类 机器视觉 HU不变矩 径向基(RBF)神经网络 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]