Ultrasonic image denoising using machine learning in point contact excitation and detection method

Area of Science:

• Materials Science
• Acoustics
• Artificial Intelligence

Background:

• Ultrasonic wave visualization in Lead Zirconate Titanate (PZT) ceramics typically employs point contact/Coulomb coupling.
• Ultrasonic signals are susceptible to various noise types (speckle, Gaussian, Poisson, salt and pepper), degrading image quality and reliability.
• Accurate structural health assessment relies heavily on high-quality ultrasonic sensor data.

Purpose of the Study:

• To implement deep learning (DL) techniques for effective noise reduction in ultrasonic images.
• To model and denoise ultrasonic images using convolutional autoencoders (CAEs).
• To quantitatively assess denoising performance using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).

Main Methods:

• Application of a point contact/Coulomb coupling technique for ultrasonic wave visualization.
• Utilizing delta pulse excitation for broadband frequency spectrum and wide directional wave vector generation.
• Implementation of deep learning-based convolutional autoencoders for noise modeling and image denoising.

Main Results:

• Convolutional autoencoders were successfully implemented for ultrasonic image denoising.
• Quantitative analysis using PSNR and SSIM metrics demonstrated the effectiveness of the DL approach.
• Speckle noise was found to perform better than other noise models based on the evaluated metrics.

Conclusions:

• Deep learning, specifically convolutional autoencoders, offers a powerful solution for denoising ultrasonic images.
• The developed method enhances the quality and reliability of ultrasonic data for structural health monitoring.
• The study highlights the potential of DL in overcoming noise limitations in ultrasonic testing applications.