基于深度学习方法的N5/NiCrAlY涂层图像识别的研究

Research on Image Recognition for NiCrAlY Coating/N5High-temperature Alloy System Based on Deep Learning Method

利用深度学习方法,将图像处理技术运用于NiCrAlY涂层/Ni基高温合金服役过程中微观形貌的图像特征信息识别和检索。以NiCrAlY涂层/N5合金为研究对象,基于获取的3600张64×64像素的截面特征图像数据集,采用深度学习技术搭建对基体的TCP相、基体与涂层的界面、氧化层这三类特征进行分类识别。分别训练有二、三层卷积层的卷积神经网络实现这三类特征的分类识别与滑动窗口检索定位。选用RMSProp优化器,配合二、三层卷积层的神经网络的测试集识别准确率分别为98%、90.67%。利用Adam优化器训练三层卷积层的卷积神经网络的测试集识别准确率为99.17%,并且此网络在检索1024×943像素图像的三大特征时表现最佳,检索正确率达到100%。

The evolution of micro-morphology for the couple of NiCrAlY coating/N5 high-temperature alloy system during service at high temperature, namely the precipitated TCP-phases within the substrate, the interface of coating/substrate, and the formed oxide scale etc., was studied by means of image processing technology, aiming to acquire the information related with their characteristics for the identification and retrieval of the relevant features of coating/alloy systems. Based on the acquired date-sets from 3600 frames of cross-sectional feature images of 64×64 pixels, a convolutional neural network(CNN) was established for classification and identification of the TCP phase, the interface of coating/substrate, and the oxide scale via a deep learning technique. The convolutional neural networks with two or three convolutional layers were respectively trained, so that the classification and identification of these three kinds of features, as well as the sliding window retrieval positioning are realized, thereby, the test set accuracy was 98% or 90.67%, respectively, for the neural network of two or three convolution layers coupled with the RMSProp optimizer. The test set accuracy for the convolutional neural network with three convolutional layers coupled with the Adam optimizer was 99.17%. This network performs best in retrieving the desired three features for images of 1024×943 pixel, correspondingly, the retrieval accuracy even can reach 100%.