Abstract: Infrared thermography has become one of the popular techniques for nondestructive testing of composite materials owing to its low cost and easy setup. However, environmental noise and inhomogeneous heating often affect thermal imaging, making it difficult to accurately identify defects in thermal images. In this paper, a hierarchical graph-based slow-feature thermography method is proposed to detect defects in composite materials. The method considers thermal image data as a time series, and constructs a multilayer hierarchical slow feature graph-based network structure, similar to a neural network, that divides the thermal image time-series into different layers of slow feature graph-based nodes independent of each other. Finally, a hierarchical graph-based slow feature analysis algorithm is applied to the graph nodes of each layer to extract local slow features for defect identification. The test results on the carbon fiber-reinforced polymer thermal image dataset show that the defect detection method based on hierarchical graph-based slow-feature thermography is feasible, and the extracted local slow features can enhance defect information. The proposed method can improve the accuracy of thermal image defect identification.