Crack detection on metallic materials fusing flying spot thermography and visible imagery with deep learning

“Flying-Spot” laser infrared thermography (FST) is a non-destructive testing technique capable of detecting small defects by scanning surfaces with a laser heat source. Defects such as cracks in metallic parts are revealed through disruptions in heat propagation, which are captured by an infrared camera. Combining this technique with visible-spectrum inspection, providing information about surface textures and geometries that are less discernible in the infrared domain, can enhance both the robustness and performance of defect detection. This work proposes the deployment of state-of-the-art visible-infrared fusion neural models for multi-spectrum crack detection on data combining visible and FST imaging. The limited amount of data available is mitigated through the deep image generation techniques for visible-FST image pairs, through the models Stable Diffusion and Control-Net. These synthetic data are used for the pre-training of multi-spectral object detection architectures for surface crack detection in metallic materials. Specifically, we assess the benefits of multi-spectral fusion compared to single-spectrum FST detection and evaluate the robustness of several state-of-the-art fusion models when handling image pairs with poor multi-spectral registration.