Automatic detection of short surface cracks in induction thermography using Deep Learning models trained with experimental data, synthetic images and FEM simulation results

Abstract

The impact of Deep Learning (DL) models in multiple areas of science and technology continues to grow at an impressive rate. Induction thermography inspection systems could greatly benefit from DL as it would allow for fully automated defect detection based on the thermal images. However, in industrial production settings the number of defective parts is usually limited, resulting in a reduced image count that hinders the training of DL models. Moreover, the annotation of the mentioned images is a time-consuming process subject to variability and human error. To address these challenges, a pipeline that combines a simpler new annotation method for short surface cracks with Generative Adversarial Networks (GANs) trained on Finite Element Method (FEM) simulations to augment the availability of images for DL model training is proposed. The performance of models trained with original data and those trained with the new pipeline is analysed using Hit/Miss Probability of Detection (POD) method. Moreover, it is demonstrated that the proposed pipeline is suitable for training DL models for the detection of short surface cracks (<3.5 mm in size) in induction thermography-based inspection systems.