Session: 05-03: Vibration Based Methods

Paper Number: 97722

97722 - Towards Computational Super-Resolution Ultrasonic Array Imaging of Material Defects via Hierarchical Multi-Scale Deep Learning 

The resolution of sensing systems is fundamentally governed by the diffraction limit, which indicates that the minimum resolvable feature size is in the order of the wavelength of a propagating wave. Imaging smaller features (e.g., hidden material defects) requires short wavelengths (or high frequencies), which could be dangerous and causes low signal-to-noise ratio due to high attenuation of the propagating wave and coherent noise due to material backscattering. Computational super-resolution sensing, aiming to recover the sub-wavelength object features (e.g., material defects) from measurements taken with insufficient wavelength, is widely pursued across many applications; however, it is ill-posed inverse problem that remains a significant challenge.

Here we present a multi-scale deep learning approach to enable super-resolution ultrasonic beamforming that computationally exceeds the diffraction limit and visualizes sub-wavelength material defects. The developed approach is a hierarchical multi-resolution deep learning framework that combines two distinct convolutional neural networks. The first network, the global detection network, globally detects the rough regions of subwavelength defects in raw low-resolution ultrasonic wave beamforming images. Subsequently, the second network, the local super-resolution network, locally resolves subwavelength-scale fine structural details of the detected defects within the detected regions. We investigate and discuss the applicability of this hierarchical multi-scale deep learning approach for computational super-resolution ultrasonic array imaging of hidden sub-wavelength defects of metallic structures.

Presenting Author: Yongchao Yang Michigan Technological University