ABSTRACT Scanning electromagnetic induction thermography (SEIT) has great potential for industrial applications, such as rapid visualising of surface cracks. This dynamic non-destructive testing method has improved detection efficiency but has led to blurred or discontinuous defect boundaries at varying scanning speeds. To alleviate this problem, we explore the use of the patch-based sparse decomposition (PSD) technique. This technique improves dynamic imaging by retaining only the information associated with sparse regression and introducing a locally adaptive threshold. Besides, this method offers a potential thermal image preconditioned scheme for follow-up artificial intelligence detection. Experimental results show that PSD can correct blurred images caused by relative motion, thus improving defect detection accuracy. Moreover, it is suitable for both translational and rotational detection systems. Images reconstructed through the deblurring process demonstrate the ability to visualise holes with a radius of 0.5 mm at velocities up to 150 mm/s, while cracks with a defect width of 0.2 mm on a railroad wheel can be detected at speeds ranging from 25 mm/s to 75 mm/s.