A cyclic self-enhancement technique for complex defect profile reconstruction based on thermographic evaluation

Although machine Learning has demonstrated exceptional applicability in thermographic inspection, precise defect reconstruction is still challenging, especially for complex defect profiles with limited defect sample diversity. Thus, this paper proposes a self-enhancement defect reconstruction technique based on cycle-consistent generative adversarial network (Cycle-GAN) that accurately characterises complex defect profiles and generates reliable artificial thermal images for dataset augmentation, enhancing defect characterisation. By using a synthetic dataset from simulation and experiments, the network overcomes the limited samples problem by learning the diversity of complex defects from finite element modelling and obtaining the thermography uncertainty patterns from practical experiments. Then, an iterative strategy with a self-enhancement capability optimises the characterisation accuracy and data generation performance. The designed loss function structure with cycle consistency and identity loss constrains the GAN’s transfer variation to guarantee augmented data quality and defect reconstruction accuracy simultaneously, while the self-enhancement results significantly improve accuracy in thermal images and defect profile reconstruction. The experimental results demonstrate the feasibility of the proposed method by attaining high accuracy with optimal loss norm for defect profile reconstruction with a Recall score over 0.92. The scalability investigation of different materials and defect types is also discussed, highlighting its capability for diverse thermography quantification and automated inspection scenarios.