Implementação de algoritmos de reconstrução de imagens por ultrassom utilizando a biblioteca WebGPU

Abstract

Motivated by the pursuit of product quality and the prevention of potential failures, ultrasonic non-destructive testing presents itself as an efficient solution, as it allows the identification of defects without altering the properties of the inspected material. With the evolution of transducers, it has become possible to collect data and reconstruct images in the B-scan format through algorithms such as SAFT, TFM, and CPWC, which offer greater accuracy in detecting discontinuities. However, the processing of these algorithms can demand high computational capacity, depending on the dataset analyzed. Although consolidated technologies such as CUDA already exist for optimizing this processing on GPUs, the present work aims to go beyond simple optimization by investigating the use of WebGPU. The goal is that, in the future, such algorithms can be incorporated into web-based applications, expanding their accessibility and applicability in industrial contexts. For the evaluation, the experiments were conducted on a laptop equipped with an NVIDIA GeForce RTX 3060 GPU, utilizing two datasets: a homogeneous aluminum block simulated in the CIVA software and real data collected from tests using the M2M Panther equipment. The results showed speed-up gains in all parallelized algorithms compared to their CPU versions. The TFM_webGPU achieved the most significant improvement, reaching a performance approximately 72 times higher due to its greater computational complexity, followed by the CPWC_webGPU with a performance about 30 times higher, and the SAFT_webGPU with a performance approximately 13.5 times higher.