Carbon fiber reinforced polymer (CFRP) composites have been widely used in aerospace, automotive and other industries, owing to their high strength and stiffness combined with low density. Thereby, there is an increased demand for defect detection in CFRP. Due to the high cost of carbon fiber, the non-destructive testing (NDT) techniques for assessment of CFPR structures have become an important research field. Pulsed thermography (PT) is a popular NDT technique for the sake of its convenient deployment and rapid detection ability, which visualizes the defects inside the CFRP specimens with thermal images. However, the thermographic data are often contaminated by measurement noise and non-uniform backgrounds, making it difficult to accurately identify defect regions. In this paper, a nonparametric signal decomposition method named multi-dimensional ensemble empirical mode decomposition (MEEMD) is utilized to decompose each thermal image into three parts, i.e. the high-frequency noise, the low-frequency backgrounds, and the signals informative for defect detection. In doing so, the noise and the non-uniform backgrounds can be removed from the thermographic data at one time, improving the detection results. The effectiveness of the proposed method is illustrated with experiments through the comparison with the existing methods.