Spatio-temporal feature extraction for infrared thermal imaging by three-dimensional convolutional autoencoder

Active infrared thermography (AIRT) combined with various machine learning methods plays an important role in the detection of internal defects in carbon fiber reinforced polymer (CFRP). However, the extraction of spatio-temporal information has not been thoroughly investigated compared to existing methods. To address this gap, this research proposes a feature extraction method called three-dimensional convolutional autoencoder thermography (3DCAT). The method fuses spatio-temporal information through 3D convolutional operations and combines the encoder’s feature extraction capability to effectively capture the potential changes and defect information in composite materials over time. Subsequently, principal component thermography was applied to the encoded data to extract the most representative components.To validate the effectiveness of three-dimensional convolution for feature extraction in thermal image data, this study visualizes and analyzes the features extracted during the training process of 3DCAT network using gradient-weighted class activation mapping. The role of 3D convolution in temporal feature extraction is revealed in depth by analyzing the degree of attention given to the temporal dimension. Ultimately, experimental results obtained from a CFRP specimen with six defects demonstrate the proposed methodology’s effectiveness.