Composite Material Defect Segmentation Using Deep Learning Models and Infrared Thermography

For non-destructive assessment, the segmentation of infrared thermographic images of carbon fiber composites is a critical task in material characterization and quality assessment. This study focuses on applying image processing techniques, particularly adaptive thresholding, alongside neural network models such as U-Net and DeepLabv3 for infrared image segmentation tasks. An experimental analysis was conducted on these networks to compare their performance in segmenting artificial defects from infrared images of a carbon-fibre reinforced polymer sample. The performance of these models was evaluated based on the F1-Score and Intersection over Union (IoU) metrics. The findings reveal that DeepLabv3 demonstrates superior results and efficiency in segmenting patterns of infrared images, achieving an F1-Score of 0.94 and an IoU of 0.74, showcasing its potential for advanced material analysis and quality control.