This paper proposes a deep learning-based system for channel estimation and signal detection in an OFDM system. This system is divided into two neural networks (NNs). The first NN is designed for semi-blind channel estimation and the second NN recovers the original signals based on channel state information (CSI) obtained from the first one. The estimation NN shows better robustness when a small number of pilots are used compared with traditional channel estimation methods. With estimated CSI, the detection NN converges within fewer epochs than the data-driven solutions. Simulation results also demonstrate the whole system works well in low SNR.