Eddy Current Pulsed Thermography (ECPT) has received much attention for its high sensitive of detectability on cracks with infrared cameras. However, when it comes to the detection in a movement way, it remains as challenges. This paper proposed an ensemble tensor decomposition to extract weak target signal of infrared thermography videos for cracks detection. The proposed algorithm jointly models the background and foreground tensor patterns as well as removing the ghosting. In order to verify the effectiveness and robustness of the proposed method, experimental studies have been carried out by applying electromagnetic thermal imaging system for cracks detection on samples with different geometry. The results of the experiments have indicated that the proposed method has significantly enhanced the contrast ratio between the defective regions and the non-defective regions.