Intelligent Ultrasonic Systems for Material Texture Recognition using Data-Efficient Neural Networks

Material texture recognition nondestructively by estimating the grain size using ultrasonic inspection has been extensively used for the characterization of material microstructures. Artificial intelligence, such as neural networks, can automate and recognize material textures accurately. However, training and deploying neural networks with high recognition accuracy is time-consuming. This limits the performance of ultrasonic applications. In this study, we investigate using several data-efficient neural networks to classify material textures using ultrasonic images. An ultrasonic testbed platform is assembled to acquire 3D ultrasonic data to train the neural networks for texture analysis. The DEUTR (Data-Efficient Ultrasonic Texture Recognition) Transformer Neural Network (TNN), is proposed to recognize material textures with high accuracy and data efficiency. The TNN utilizes a simple network architecture that replaces the convolutions and recurrence entirely with the attention mechanism resulting in reducing training and execution time. For performance comparison, several data-efficient Convolutional Neural Networks (CNN): NasNetMobile, MobileNet, Xception, were trained with Transfer Learning (TL) to learn grain scattering features using ultrasonic images. By training these neural networks, we obtained the average training and testing accuracy of 99.47% and 98.04% to recognize material textures and the highest image throughput of 148 images/seconds using the DEUTR transformer neural network on testing. We aim to build an intelligent ultrasonic system to recognize material textures with high accuracy at the microstructural level for data-efficient NDE applications.