DAE-SWnet: Unsupervised internal defect segmentation through infrared thermography with scarce samples

Abstract

Accurate defect detection is essential for manufacturing reliability, yet identifying internal defects remains challenging. While infrared thermography offers distinct advantages for internal inspection, its effectiveness is hindered by noise and uneven heating. Traditional image processing algorithms struggle with these nonlinearities, while supervised deep learning methods require large annotated datasets, usually impractical in industrial settings. To overcome data scarcity, we propose a novel infrared data amplification strategy leveraging the dynamic temperature evolution. By varying defect depth and heating duration, we generate 16000 thermal images from merely two defective samples using pulsed thermography. Furthermore, we introduce an unsupervised defect segmentation framework, Deep Autoencoder with Swin Transformer Wnet (DAE-SWnet). First, a deep autoencoder denoises thermal images, during which we discover a non-monotonic relationship between reconstruction loss and denoising performance. Next, Swin Transformer and Wnet are cohesively integrated with optimized encoder-decoder channels, to extract latent defect features from denoised images. These latent representations are postprocessed to obtain final segmentation results. Trained solely on artificially designed defects, our model exhibits exceptional performance across samples with varying materials and defect shapes. Moreover, comparative experiments demonstrate that the model achieves higher precision, stronger stability, and better generalization to real-world manufacturing processes. Specifically, it achieves 74.0 % IoU, 84.3 % F1-score, and 97.7 % accuracy, with an average inference time of 0.278 s, highlighting its superiority and practical potential in industrial scenarios where defective products are difficult to obtain and annotate.