We propose our own convolutional neural network (CNN) structure as a post-processing method for the rapid THz Imaging in Terahertz Time Domain Spectroscopy(THz-TDS). The experiment results show that our approach can greatly reduce the noise and artifacts in the collected under-sampled THz images, which in turn benefits the rapid THz imaging technique by allowing higher acquisition rates. In our setup, 5 times higher acquisition rates can be realized. Moreover, this approach needs no extra hardware cost, and since the training has been done offline, the processing time in practice can be ignored. To the best of our knowledge, this is the first time applying deep learning(DL) methods into rapid THz imaging technique, which would inspire researchers to explore more DL based applications and technologies for THz technique development.