Improvement of defect intelligent recognition of laser thermography with enhanced virtual dataset

The present study proposes a hybrid framework based on the combination of Convolutional Neural Network (CNN) and Transformer for the improvement of defect identification in laser infrared thermography testing (IRT) technology. This framework utilises virtual datasets generated through simulation to train the model, enabling accurate prediction of real defects. The initial stage of the framework involves a time-domain noise reduction convolutional neural network model based on Generative Adversarial Networks (GAN). The subsequent stage incorporates a defect recognition model that adopts a hybrid structure of CNN and Transformer to discern disparities between spatial noises and defect signals. The present study employs a limited amount of simulation data for data augmentation and perturbation, thereby generating a virtual dataset to be used as training material for model development. Experiment results show that the model proposed exhibits much higher performance even with model parameters less more than 50%, allowing for fast real-time inference or deployment on less powerful devices.