Deep learning based segmentation using full wavefield processing for delamination identification: A comparative study

In this paper, several deep fully convolutional neural networks for image segmentation such as residual UNet, VGG16 encoder-decoder, FCN-DenseNet, PSPNet, and GCN are employed for delamination detection and localisation in composite materials. All models were trained and validated on our previously generated dataset that resembles full wavefield measurements acquired by scanning laser Doppler vibrometer. Additionally, a thorough comparison between all presented models is provided based on several evaluation metrics. Furthermore, the models were verified on experimentally acquired data with a Teflon insert representing delamination showing that the developed models can be used for delamination size estimation. The achieved accuracy in the current implemented models surpasses the accuracy of previous models with an improvement up to 22.47% for delamination identification.

Automated summary

This paper investigates several deep fully convolutional neural network models for delamination detection and localization in composite materials. The models were trained and validated on a specific dataset and achieved improved accuracy compared to previous models.