Infrared thermography (IRT) stands out as one of the most reliable and efficient techniques for identifying delamination inside concrete structures due to its accuracy among nondestructive evaluation methods in the structural health monitoring field. However, the traditional IRT approach is laborious and time-consuming, even prone to risk for investigators. Applying machine learning to IRT images is a promising alternative to automatic recognition of delaminations of concrete structures, which can address the limitations of conventional IRT investigation. Unfortunately, this approach has received much less attention, primarily because of the lack of IRT data. Hence, this study attempted to utilize several advanced machine learning algorithms on a self-prepared IRT dataset to classify the IRT images into defective or non-defective categories. The efficiency of selected machine learning models was comprehensively investigated, and the most effective algorithms for hidden defect classification were proposed. Moreover, the data augmentation was applied to the selected models, which examined its effectiveness in enhancing the accuracy of machine learning-based classifiers on the IRT dataset.