One of the main barriers in front of Free Space Optical (FSO) communication systems is the atmospheric turbulence induced fading. Theoretically, the Maximum Likelihood (ML) detector is the optimum detector. The ML detector requires Channel State Information (CSI), which can be provided in perfect or blind forms. The perfect CSI ML detector requires pilot transmission for channel estimation, which increases the complexity and reduces the data rate The blind CSI ML detector uses blind channel estimation, which leads to performance degradation. In this paper, for the first time, an efficient and low complexity deep learning based detector is presented for FSO system. The proposed deep learning based detector does not require CSI at all, it feeds the received signal directly into a deep neural network. The proposed deep learning based detector is compared with perfect CSI ML detector and blind CSI ML detector. In this paper, log-normal, gamma-gamma, and negative exponential distributions are considered for modeling weak, weak to strong, and saturate atmospheric turbulence regimes, respectively. Results indicate that the performance of proposed deep learning based detector gets close enough to the perfect CSI ML detector, with a significantly lower complexity than the blind CSI ML detector. The proposed detector is almost 80 times faster than blind CSI ML detector. In addition, it does not have an error floor, while one of the main problems of blind CSI ML detector is the error floor. Besides much less complexity, the proposed detector has almost the same performance as blind CSI ML detector at weak atmospheric turbulence regime. The available blind CSI ML detectors are practical only in weak turbulence, because they assume that channel coefficients are constant for the duration of some symbols. However, the proposed deep learning based detector does not consider this assumption, and can be used in all atmospheric turbulence regimes. The performance of the proposed detector degrades when atmospheric turbulence gets stronger. For instance, the performance of the proposed deep learning based detector degrades 7 dB compared with blind CSI ML detector at target bit error rate of 10−3. However, the proposed deep learning based detector outperforms blind CSI ML detector at high signal to noise ratios, because in this range blind CSI ML detector suffers from the error floor.

Keywords