Infrared Thermographic Data Processing with Deep Learning and Explainable AI

Pulsed thermography (PT), as an important nondestructive testing method, has attracted increasing attention in recent years, due to its rapid measurement speed, excellent mobility, and high resolution. Its applicability spans across various materials, such as metal and composite materials. The thermographic data consist of two-dimensional infrared sequences representing the evolution of the surface temperature of the test specimen. Even for skilled professionals, it is challenging to accurately identify all kinds of defects. Furthermore, the characterization for the defects can be even more difficult to decide. Although deep learning has become a popular automated data processing approach in recent years, there are some common issues when they are applied to the analysis of thermographic image sequences. First, deep learning approaches are typically data-hungry, whereas thermographic data are scarce as the preparation for testing is usually tedious. Second, as PT focuses on the detection of subsurface defects, it is often quite challenging to obtain the exact location of the defects. However, the accurate characterization of the defects is crucial for the training of deep neural networks. Third, unlike the computer vision field, where there are numerous well-prepared datasets for benchmarking different algorithms, there are only a few such publicly accessible PT datasets. However, these datasets are fundamental for advancing algorithms in thermographic data processing. Fourth, the deep learning models, despite their good performance, are often considered "black boxes". This presents an obstacle to their deployment in the industry for quality assurance due to the significant challenge of learning the trust of the end user who bears the responsibility for the product’s quality. This study investigates the aforementioned issues. To address the scarcity of data and the limited benchmark datasets, two PT datasets are prepared and made publicly available, one is for defect segmentation and the other is for defect depth estimation and localization. The latter dataset is labeled using the corresponding CAD images. This enhances the accuracy of the labeling information. Furthermore, to enhance the explainability of the deep models used for infrared data processing, three explainable AI methods are investigated, including the activation map visualization method, feature attribution method, and feature occlusion method. The activation map visualization method shows the decision-making process of the deep model, which is similar to human perception. The feature attribution method and feature occlusion method generated similar heat maps, demonstrating that the models use the correct features to make final decisions. The trust crisis of the deep model can therefore be mitigated.