Employing a U-net convolutional neural network for segmenting impact damages in optical lock-in thermography images of CFRP plates

Carbon fibre reinforced plastics (CFRPs) are replacing metals in fields such as aerospace due to their high mechanical strength and low weight. They have an anisotropic behaviour, which hinders the analysis of structural impairment caused by damages like impacts. Optical lock-in thermography (OLT) can be used to assess CFRP integrity and image processing tools can be applied to measure the area affected by impacts on the thermal images. There are several alternatives for segmenting those images and this work proposes a transfer learning approach with a U-Net neural network used in characterisations of neuronal structures in microscopy for segmenting OLT images of CFRP plates with impact damages. After training and testing this tool with OLT images, using as ground truth their manual segmentation, the results were compared with four image processing combinations of methods: a filter based on two-dimensional Fast Fourier Transform with an adaptive threshold tool; an absolute thermal contrast (ATC) with a global threshold (GT) tool; the image overflow difference with GT; and principal component analysis (PCA) with GT. The results show that the U-Net was the most reliable for the proposed conditions for defective area assessment, allowing a higher safety in maintenance tasks.