Comparison between segmentation performances of a tool based on wavelet decomposition and multithreshold and of a U-net convolutional neural network applied to shearography images of carbon fiber reinforced plastic plates with low-velocity impact damages

Carbon fiber reinforced plastics (CFRPs) have been used to replace metallic alloys in many industries because of their high strength-to-weight ratio. Due to their anisotropic behavior, low-velocity impacts can produce defects whose effects on material performance are hard to foresee. Nondestructive testing (NDT) methods are a convenient alternative to evaluate their integrity. Shearography, an image-based optical interferometric technique for measuring deformation, is one of these NDT possibilities. The segmentation of defects in the resulting images provided by such a method is essential to correctly locate and indicate the severity of impact damage. This task is especially intricate for shearography images with barely visible impact damage because of their usual low signal-to-noise ratio. We compare a combination of wavelet decomposition with multithresholds introduced in a previous publication with a U-net convolutional neural network for analyzing impact damage in CFRP plates. Both tools are detailed and then evaluated using the Matthews correlation coefficient and the equivalent diameter criterion. The results showed that U-net provided a better impact damage characterization in both evaluation metrics, allowing a safer defect detection that is less dependent on the inspector's ability to interpret them.