侯劲尧, 刘卫国, 周顺, 高爱华, 葛少博, 肖相国. 基于卷积神经网络的光学元件表面缺陷图像分类[J]. 应用光学, 2023, 44(3): 677-683. DOI: 10.5768/JAO202344.0305003
引用本文: 侯劲尧, 刘卫国, 周顺, 高爱华, 葛少博, 肖相国. 基于卷积神经网络的光学元件表面缺陷图像分类[J]. 应用光学, 2023, 44(3): 677-683. DOI: 10.5768/JAO202344.0305003
HOU Jinyao, LIU Weiguo, ZHOU Shun, GAO Aihua, GE Shaobo, XIAO Xiangguo. Image classification of optical element surface defects based on convolutional neural network[J]. Journal of Applied Optics, 2023, 44(3): 677-683. DOI: 10.5768/JAO202344.0305003
Citation: HOU Jinyao, LIU Weiguo, ZHOU Shun, GAO Aihua, GE Shaobo, XIAO Xiangguo. Image classification of optical element surface defects based on convolutional neural network[J]. Journal of Applied Optics, 2023, 44(3): 677-683. DOI: 10.5768/JAO202344.0305003

基于卷积神经网络的光学元件表面缺陷图像分类

Image classification of optical element surface defects based on convolutional neural network

  • 摘要: 光学元件的表面疵病,即表面缺陷,其形状的大小会直接影响光学系统的性能,在对表面缺陷进行分类时,所面对的很多表面缺陷的形状都是不规则的,依靠普通的模式识别技术,分类很难达到预期的效果。为解决精密光学元件表面缺陷分类方法中精度低、耗时长的问题,提出了基于卷积神经网络的精密光学元件表面缺陷分类方法。采用散射法获取表面缺陷图像,分析其成像特点,通过对图像进行旋转,镜像扩增了数据集,加强了网络的训练能力。使用AC训练网络模型,在不增加额外计算量的同时加强了网络的特征获取力。通过Softmax分类器,将精密光学元件表面缺陷分为划痕、麻点及噪点3类。实验结果表明,所使用的模型对缺陷分类精度超过99.05%。

     

    Abstract: The surface defects of optical elements, namely surface defects, will directly affect the performance of the optical system. In the classification of surface defects, the shapes of many surface defects are irregular, so it is difficult to achieve the expected effect by relying on normal pattern recognition technology. To overcome the low precision and long time consuming in classification of surface defects of precision optical elements, a classification method of surface defects based on convolutional neural network was proposed. Firstly, the surface defect image was obtained by scattering method to analyze its imaging characteristics, and the training ability of the network was strengthened by rotating the image and mirroring the amplified dataset. Furthermore, the AC training network model was used to strengthen the feature acquisition ability of the network without increasing the extra calculation. Finally, the Softmax classifier was used to classify the surface defects into scratch, pitting and noise. The experimental results show that the defect classification accuracy of the used model is more than 99.05%.

     

/

返回文章
返回