By considering the influences of atmospheric turbulence which leads to performance degradation of the maximum likelihood (ML) detector for the quadrature amplitude modulation (QAM) signal .In this paper, a signal detector based on deep-learning neural network (DNN) is presented for ACO-OFDM free-space optical communication (FSO) systems.The proposed DNN detector with a fully connecte d layer realized the channel-blind estimation,channel equalization,and signal demodulation in the system without pilot.Results indicate that the performance of DNN detector is better,which is trained at moderate-strong turbulence regime,and the performance remains robustness when atmospheric turbulence gets stronger.For instance the bit error rate of 8QAM,16QAM and 64QA M decrease to around 2×10－5,5×10－5 and 5×10－4 respectiv ely.