Photoelastic and Stokes images through deep convolutional neural networks: a comparison of stress fields

Abstract

Photoelasticity is a non-destructive optical testing technique that focuses on stress analysis. Traditional methods of demodulating stress fields are limited by various conditions, such as the image acquisition set, material properties, load values, light sources and isoclinics. As an alternative, deep convolutional neural networks (DCNNs) have been used to recover stress fields in automated and predictive methods. In this study, different DCNNs architectures are trained by means of two datasets, each one with 45000 images. First dataset has images with four polarization states (0°, 45°, 90° and 135°). Second dataset has images with 3-channel, each one corresponding to a Stokes parameter (s0, s1, s2). The quality of predicted images is evaluated with quality metrics such as MSE, SSIM, and PSNR. MSE and Adam are used as loss function and optimizer, respectively. Results show that on average, the use of DCNNs with images with four polarization states achieve better quality metrics than DCNNs with Stokes images. These results indicate that it is possible to obtain real-time stress fields using different representations of polarized images in deep networks and opens new opportunities for representing polarized images in deep learning models and extending its applications of stress analysis.