A multiscale residual U-net architecture for super-resolution ultrasonic phased array imaging from full matrix...

Area of Science:

• Ultrasonic phased array imaging
• Deep learning applications in materials science
• Nondestructive testing (NDT)

Background:

• Full-matrix capture (FMC) in ultrasonic imaging utilizes extensive echo data for improved visualization.
• Conventional beamforming methods like delay-and-sum have limitations in resolution and contrast-to-noise ratio for FMC data.
• Addressing these limitations is crucial for advancing ultrasonic imaging capabilities.

Purpose of the Study:

• To introduce a deep learning (DL)-based image formation method, FMC-Net, for reconstructing high-quality ultrasonic images directly from FMC data.
• To leverage DL's nonlinear mapping capabilities to model complex wave-matter interactions.
• To enhance the resolution and visualization performance beyond conventional ultrasonic imaging techniques.

Main Methods:

• Developed FMC-Net, an encoder-decoder DL architecture utilizing multiscale residual modules.
• Trained the network end-to-end to map FMC data to the acoustic scattering coefficient distribution.
• Employed numerical simulations and experimental validation for performance assessment.

Main Results:

• FMC-Net demonstrated superior performance compared to the total focusing method and wavenumber algorithm.
• The DL approach achieved enhanced resolution, exceeding conventional limits.
• Subwavelength defects were effectively visualized, showcasing improved imaging fidelity.

Conclusions:

• The proposed FMC-Net offers a powerful DL-based alternative for ultrasonic image formation from FMC data.
• This approach significantly outperforms traditional methods in resolution and defect visualization.
• FMC-Net holds promise for diverse applications in ultrasonic array imaging and NDT.