摘要
针对小样本条件下卷积神经网络过拟合导致航空叶片表面缺陷检测精度不高的问题,提出了一种集合辅助分类器和条件生成对抗网络(CGAN)的数据增强方法以提高叶片表面缺陷检测精度。利用双边滤波和Laplacian高通滤波对图像进行去噪和增强,通过随机裁剪和仿射变换预增强数据集,将随机噪声z和条件y输入生成器,训练网络模型的参数,使用训练好的生成器模型生成符合真实样本分布的增强数据。实验结果表明,该方法对比仿射变换方法,卷积神经网络的分类性能得到了显著地提升,叶片表面缺陷检测准确率提高了8.3%,达到94.5%。
For small samples,in view of the problem that the over-fitting of the convolutional neural network leads to the low accuracy of the surface defect detection of aviation blades,a data augmentation method that combine a auxiliary classifier with conditional generative adversarial network(CGAN)were proposed to improve the blade surface defect detection accuracy.Bilateral filtering and Laplacian high-pass filtering were used to denoise and enhance the image,the data set was pre-enhanced through random cropping and affine transformation,input random noise vector z and condition vector y into the generator,trained the parameters of the network model,the trained generator model was used to generate augmentation data that conforms to the real sample distribution.Experimental results show that the classification performance of the convolutional neural network is significantly improved by this method compared with the affine transformation method.The detection accuracy of blade surface defects is increased by 8.3%and reaches 94.5%.
作者
丁鹏
卢文壮
刘杰
袁志响
DING Peng;LU Wen-zhuang;LIU Jie;YUAN Zhi-xiang(College of Mechanical&Electrical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China)
出处
《组合机床与自动化加工技术》
北大核心
2022年第7期18-21,共4页
Modular Machine Tool & Automatic Manufacturing Technique
基金
国家自然科学基金资助项目(51975287)
南京航空航天大学创新计划项目(kfjj20200502)。
关键词
叶片
缺陷检测
数据增强
图像处理
生成对抗网络
blade
defect detection
data augmentation
image processing
generative adversarial network(GAN)
