Visual geometry Group-UNet: Deep learning ultrasonic image reconstruction for curved parts

Abstract

Detecting small defects in curved parts through classical monostatic pulse-echo ultrasonic imaging is known to be a challenge. Hence, a robot-assisted ultrasonic testing system with the track-scan imaging method is studied to improve the detecting coverage and contrast of ultrasonic images. To further improve the image resolution, we propose a visual geometry group-UNet (VGG-UNet) deep learning network to optimize the ultrasonic images reconstructed by the track-scan imaging method. The VGG-UNet uses VGG to extract advanced information from ultrasonic images and takes advantage of UNet for small dataset segmentation. A comparison of the reconstructed images on the simulation dataset with ground truth reveals that the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) can reach 39 dB and 0.99, respectively. Meanwhile, the trained network is also robust against the noise and environmental factors according to experimental results. The experiments indicate that the PSNR and SSIM can reach 32 dB and 0.99, respectively. The resolution of ultrasonic images reconstructed by track-scan imaging method is increased approximately 10 times. All the results verify that the proposed method can improve the resolution of reconstructed ultrasonic images with high computation efficiency.