奥氏体不锈钢焊缝超声相控阵检测图像智能分类

Image Intelligent Classification of Ultrasonic Phased Array Testing for Austenitic Stainless Steel Weld

奥氏不锈钢焊缝的粗晶粒以及材料各向异性,导致超声检测图像信噪比较低,人工检测判断缺陷存在困难。分别使用InceptionV3、ResNet50、DenseNet121和MobileNetV2卷积神经网络来进行奥氏不锈钢焊缝相控阵B扫图像的自动分类。实验结果表明,大型卷积神经网络分类准确率高,ResNet50和DenseNet121在测试集上的分类准确度达到100%,Inception V3达99.8%,但轻量化的卷积神经网络MobileNetV2的准确率只有90.8%,存在严重过拟合。通过改进轻量化卷积神经网络MobileNetV2的训练方式,并引入Dropout层,可以保证分类准确度达到100%,并且减小了大量网络参数,更利于移动端的应用。

The coarse grain and material anisotropy of austenitic stainless steel weld lead to low signal-to-noise ratio of ultrasonic inspection image, and it is difficult to judge defects by manual inspection. Firstly, InceptionV3, ResNet50, DenseNet121 and MobileNetV2 convolutional neural networks are used to automatically classify phased array B-scan images of austenitic stainless steel welds. The experimental results show that the classification accuracy of large convolutional neural network is high. The classification accuracy of ResNet50 and DenseNet121 on the test set is 100%, and the classification accuracy of InceptionV3 is 99.8%. However, the accuracy of lightweight convolutional neural network MobileNetV2 is only 90.8%, which has serious over fitting;Further, by improving the training method of lightweight convolutional neural network MobileNetV2 and introducing Dropout layer, the classification accuracy can reach 100%, and a large number of network parameters can be reduced, which is more conducive to the application of mobile terminal.