Masked Sequence Autoencoding for Enhanced Defect Visualization in Active Infrared Thermography

TL;DR

This paper introduces a masked sequence autoencoding framework with attention mechanisms for active infrared thermography, significantly improving defect visualization and detection accuracy in non-destructive testing.

Contribution

It proposes a novel Masked CNN-Attention Autoencoder that captures local and global thermal features, enabling efficient training and better defect analysis under varied conditions.

Findings

Outperforms existing autoencoders in contrast and SNR

Reduces training time by a factor of 30

Effective across different material specimens

Abstract

Active infrared thermography (AIRT) became a crucial tool in aerospace non-destructive testing (NDT), enabling the detection of hidden defects and anomalies in materials by capturing thermal responses over time. In AIRT, autoencoders are widely used to enhance defect detection by reducing the dimensionality of thermal data and improving the signal-to-noise ratio. However, traditional AIRT autoencoders often struggle to disentangle subtle defect features from dominant background responses, leading to suboptimal defect analysis under varying material and inspection conditions. To overcome this challenge, this work proposes a Masked CNN-Attention Autoencoder (AIRT-Masked-CAAE) that integrates convolutional feature extraction with attention mechanisms to capture both local thermal patterns and global contextual dependencies. The AIRT-Masked-CAAE introduces a masked sequence autoencoding strategy, where the network learns to infer missing thermal responses from surrounding contextual cues, while suppressing background redundancy. In addition, the proposed masked sequence autoencoding approach enables training on only a subset of the thermal sequence, while providing generalizable latent representations and reducing training time by a factor of 30. The AIRT-Masked-CAAE framework was evaluated using specimens made of PVC, CFRP, and PLA. The results demonstrate that the AIRT-Masked-CAAE surpasses state-of-the-art AIRT autoencoders in terms of contrast, signal-to-noise ratio (SNR), and metrics based on neural networks.