Thermal Inspection Dataset for Defect Segmentation in CFRP Laminates
Description
This dataset contains thermal image data and corresponding annotations from a pulsed thermography (PT) inspection conducted on a carbon fiber-reinforced polymer (CFRP) laminate. The dataset is designed to support research in defect segmentation, early defect detection, and related applications in composite material inspections. Dataset Composition: Images: A total of 1,034 thermal images captured using a midwave infrared (MWIR) camera at a resolution of 640 × 512 pixels and a frame rate of 55 Hz. These images were recorded during PT inspections following a short, intense heat pulse. Annotations: Each of the thermal images has a corresponding expert-annotated image, resulting in 1,034 segmentation masks. The annotations were created using the VGG Image Annotator, a tool developed by the Visual Geometry Group at the University of Oxford, and represent the ground truth data for model training and validation. Sample Details: Material: The sample is a unidirectional carbon/PEEK laminate with a fiber volume fraction of 61%. The stacking lay-up of the laminate is [02/902]6, and its dimensions are 100 × 100 mm. Defect Characteristics: Artificial defects in the form of Kapton tape inserts were introduced into the laminate during the moulding process. These defects vary in size and depth: Nominal Depths: D1: 0.13 mm D2: 0.26 mm D3: 0.39 mm Nominal Sizes: 2 × 2 mm 3 × 3 mm 4 × 4 mm Inspection Methodology: A MWIR camera captured thermal profiles as the surface of the laminate cooled after heat application. Images display the temporal evolution of surface defects, highlighting changes over time. This temporal variation is critical for identifying and segmenting defects accurately. Applications: Defect segmentation in composite materials. Development and validation of machine learning models for thermal image analysis. Early detection and tracking of defect progression. Ground Truth and Segmentation Masks: The annotated dataset includes: Segmentation masks created from the annotations, used for model training and validation. If you use this dataset in your research, please acknowledge the authors and cite the associated publication. For more details or to access the dataset, please contact the authors. Associated publication: Garcia Vargas, I., & Fernandes, H. (2025). Spatial and temporal deep learning algorithms for defect segmentation in infrared thermographic imaging of carbon fibre-reinforced polymers. Nondestructive Testing and Evaluation, 1–21. https://doi.org/10.1080/10589759.2025.2457593
Files
Steps to reproduce
Data Acquisition and Reproducibility This dataset was created through a systematic process involving material preparation, defect introduction, pulsed thermography (PT) inspection, and image annotation. Below is a concise summary of the methods, tools, and workflows used, aimed at ensuring reproducibility. 1. Material Preparation The sample is a unidirectional carbon/PEEK laminate with a fiber volume fraction of 61% and a stacking lay-up of [02/902]6. Its dimensions are 100 × 100 mm. Artificial defects were introduced using Kapton tape inserts with nominal sizes of 2 × 2 mm, 3 × 3 mm, and 4 × 4 mm. These were placed at specific depths (D1: 0.13 mm, D2: 0.26 mm, and D3: 0.39 mm) in different layers prior to the moulding process to simulate internal defects. 2. Inspection Protocol Pulsed thermography was used as the inspection technique. A short, intense pulse of heat was applied to the laminate's surface, and a midwave infrared (MWIR) camera recorded thermal profiles over time. The camera captured images at a resolution of 640 × 512 pixels with a frame rate of 55 Hz, producing 1,034 thermal images. These images display the temporal evolution of defects, which is critical for segmentation and defect detection research. 3. Annotation Process The images were annotated using the VGG Image Annotator (VIA). An expert manually labeled the defect regions, creating segmentation masks that serve as the ground truth dataset. Each annotated mask corresponds directly to an original thermal image, providing a comprehensive dataset for training and validation. 4. Tools and Instruments Kapton Tape: Used to create artificial defects. MWIR Camera: For capturing thermal images during the PT inspection. Heat Source: Delivered a pulsed thermal stimulus to the sample surface. VGG Image Annotator: Used for manual defect annotation. 5. Workflow Summary The laminate was prepared with embedded defects during the moulding process. Pulsed thermography was conducted to capture thermal images. Thermal images were annotated using the VIA tool. The dataset was compiled, including original images and corresponding segmentation masks. 6. Reproducibility Notes Reproducing this dataset requires similar materials, defect introduction methods, and pulsed thermography conditions. The MWIR camera settings and the annotation process should align closely with the described protocols. By following these steps, researchers can replicate or extend the study, contributing to advancements in defect detection and segmentation.