Deep Convolutional Neural Networks Applied to Ultrasonic Images for Material Texture Recognition

Texture recognition nondestructively by estimating the grain size has been widely used for the characterization of the physical and structural integrity of materials. As the ultrasonic signal passes through the materials, signal energy attenuates due to scattering and absorption, which are functions of the frequency and grain size distribution. Thus, the scattering and attenuation of ultrasonic echoes can be utilized for grain size evaluation and microscopic texture analysis. In this paper, we propose to investigate the performance of using deep convolutional neural networks (CNNs) to learn grain scattering features and classify materials. An ultrasonic testbed platform is assembled to obtain 3D ultrasonic data from heat-treated steel blocks with different grain sizes. The 3D acquired data are utilized to construct 2D images (B-Scans and C-Scans) to train the proposed deep CNNs classifiers for texture analysis. Several state-of-the-art deep CNNs are trained and compared to classify the grain scattering textures of three heat-treated steel blocks. These deep CNN classifiers are pre-trained on large datasets (ImageNet) followed by further training with transfer learning (TL) using experimental ultrasonic images. A lightweight TL based deep CNN classifier known as LightWeightTextureNet (LWTNet) was utilized to classify material textures with high validation accuracy of 99.58%.