TriMap thermography with convolutional autoencoder for enhanced defect detection of polymer composites

Abstract

Pulsed thermography data are typically affected by noise and uneven backgrounds, thereby complicating defect identification. Hence, various image analysis methods have been applied to improve defect detectability. However, most of them directly analyze the original images, while the low quality of the data is disregarded. Herein, a thermographic data analysis method named TriMap thermography with convolutional autoencoder (CAE) is proposed to overcome this problem. In this method, a CAE is used to reduce noise and enhance the quality of thermograms. Subsequently, the TriMap algorithm is used to extract features from the enhanced data. Specifically, the TriMap uses triplet information to improve the low-dimensional embedding quality and obtain an abstract representation of high-dimensional data. Finally, defects and uneven backgrounds are effectively distinguished by visualizing the embedding vectors. The test results of a carbon fiber-reinforced polymer specimen validate the effectiveness of the proposed method.