摘要
为利用有限的钛合金烧伤图像来提高图像识别准确率,设计一种基于迁移学习的钛合金烧伤图像识别方法。利用该方法将经过ImageNet数据集训练的GoogLeNet和ResNet50网络模型保留卷积层及其对应权重参数作为特征提取器,分别与设计的特征识别网络建立连接,形成烧伤图像识别网络。在自制钛合金烧伤图像数据集上对网络模型进行训练。通过实验验证表明:该模型可以自动识别出无烧伤、轻度烧伤、中度烧伤、重度烧伤4类钛合金图像,识别准确率较未使用迁移学习的网络模型有大幅提升。
To improve the accuracy of image recognition with limited titanium alloy burn images, a titanium alloy burn image recognition method based on transfer learning is designed, by which, GoogLeNet and ResNet50 network models trained on the ImageNet data set retain the convolutional layer and its corresponding weight parameters are treated as feature extractors to establish connections with the designed feature recognition network respectively so as to form a burn image recognition network. The network model on the self-made titanium alloy burn image dataset is trained. The experimental verification shows that the model can automatically identify four types of titanium alloy images with non-burns, mild burns, medium burns, and severe burns, and the recognition accuracy is greatly improved compared with the network model without transfer learning.
作者
周光辉
卢文壮
张其真
丁鹏
吴泊鋆
ZHOU Guanghui;LU Wenzhuang;ZHANG Qizhen;DING Peng;WU Bojun(College of Mechanical and Electical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China)
出处
《机械制造与自动化》
2022年第2期111-115,共5页
Machine Building & Automation
关键词
迁移学习
图像识别
钛合金
磨削烧伤
transfer learning
image recognition
titanium alloy
grinding burn