Underwater wireless optical communication has garnered significant attention in the wireless communication field due to its high data rate, enhanced security, and lightweight nature. However, seawater can induce absorption and scattering of light. Absorption results in a reduction of the received optical power at the receiver, which is an irreversible process, while scattering causes alterations in the received photons at the receiver. Moreover, the ocean typically contains turbulence, a phenomenon caused by temperature variations and irregular movements, leading to random fluctuations in the optical signal. Consequently, the underwater channel is intricate and challenging to predict. To achieve reliable communication performance, a more dependable signal detection method is required at the receiver. In this study, a deep learning-assisted signal detection method is proposed for underwater optical communication. A convolutional neural network (a specialized form of deep neural network) is developed to directly detect the Original On-off Keying (OOK) signal, and two distinct training methods for the Deep Neural Network (DNN) are proposed during the training phase. Initially, an indoor underwater optical communication experimental platform is designed and constructed, incorporating three types of water tank channels (flowing water, turbid flow 1, turbid flow 2). The attenuation coefficients and probability density functions of the channels are measured. Subsequently, a simulated underwater optical channel is derived based on the measured channel mathematical models, and a simulated dataset of OOK signals for the neural network is obtained. The proposed methods are tested using the dataset, and the performance of the two different DNN training methods and the adaptive threshold method is simulated under different simulated channels. The proposed methods exhibit an improvement in Bit Error Rate (BER) compared to the adaptive threshold method at any signal-to-noise ratio in the three channels. The improvement is most notable in the simplest flow channel, with up to a two-order-of-magnitude enhancement, and it increases with higher signal-to-noise ratios in the relatively complex turbid flow channel 2. Additionally, due to DNN training method 1 learning multiple datasets from different channels, it exhibits worse BER performance compared to training method 2, which only learns one channel dataset. However, thanks to the powerful fitting capability of DNN, the BER is still superior to the adaptive threshold method. To validate the simulation results, experimental datasets of OOK signals are obtained based on the experimental platform. The DNN is retrained and tested using the experimental datasets, and the BER performance of the two different DNN training methods and the adaptive threshold method is experimentally studied. For 5 Mbps communication transmission in the three water tank channels, the DNN method achieves a reduction in BER of two orders of magnitude, one order of magnitude, and one order of magnitude, respectively, compared to the adaptive threshold method. The trend of the experimental results is consistent with the simulation. For turbid flow 1 and different communication rates (5 Mbps, 10 Mbps, 25 Mbps), the DNN method achieves a reduction in BER of one order of magnitude at all three rates, and the proposed method requires lower received optical power compared to the adaptive threshold method when the BER is the same. The simulation and experimental results demonstrate that the proposed method enhances the performance of underwater wireless optical communication in complex channels compared to the adaptive threshold method, validating the reliability of the method. Therefore, the method can offer valuable insights for the design of high-speed and reliable underwater wireless optical communication systems.