Here we propose a new scheme to reconstruct the baryon acoustic oscillations (BAO) signal, with key cosmological information, based on deep convolutional neural networks. After training the network with almost no fine-tuning, in the test set, the network recovers large-scale modes accurately: the correlation coefficient between the ground truth and recovered initial conditions still reach $90\%$ at $k \leq 0.2~ h\mathrm{Mpc}^{-1}$, which significantly improves the BAO signal-to-noise ratio until the scale $k=0.4~ h\mathrm{Mpc}^{-1}$. Furthermore, our scheme is independent of the survey boundary since it reconstructs initial condition based on local density distribution in configuration space, which means that we can gain more information from the whole survey space. Finally, we found our trained network is not sensitive to the cosmological parameters and works very well in those cosmologies close to that of our training set. This new scheme will possibly help us dig out more information from the current, on-going and future galaxy surveys.