Material Texture Recognition using Ultrasonic Images with Transformer Neural Networks

Material texture recognition by estimating the grain size has been extensively used for characterization of material structures. Ultrasonic inspection can approximate material grain size nondestructively with advantages of one-sided measurement, high penetration depth and inspection accuracy. In ultrasonic testing, the energy of signal attenuates as ultrasonic signal propagates through the material. This attenuation is due to scattering and absorption, which are functions of the frequency and grain size distribution. Therefore, the attenuation and scattering of ultrasonic echoes can be used to evaluate grain size for microscopic texture. In this paper we propose to use the transformer neural networks to learn grain scattering features for material textures recognition. The transformer neural network utilizes the multi-head attention mechanism to substantially reduce the computation cost. An ultrasonic testbed platform is assembled to acquire the 3D ultrasonic data cube to train the neural networks for texture analysis. The 3D data cube consists of a sequence of 2D ultrasonic C-scan images and is obtained from three different heat-treated steel blocks. Several state-of-the-art machine learning algorithms, the deep Convolutional Neural Networks (CNNs) and Support Vector Machine (SVM) were trained and compared to classify the grain scattering textures of three heat-treated steel blocks. To build a data-efficient automatic system for ultrasonic nondestructive evaluation (NDE) applications, a self-attention based transformer neural networks: Ultrasonic Texture Recognition Vision Transformer: UTRV Transformer, was proposed to classify material textures with high testing accuracy of 96.15%.