Interpretation of seismic surveys is severely constrained until subsurface velocity information is extracted by seismic survey interpreters. Velocity modeling is an important aspect of seismicity, and the extraction of accurate velocities of subsurface media is an important parameter for obtaining high-precision imaging. Conventional velocity information can be obtained by layer inversion and full waveform inversion (FWI), but the conventional methods are computationally intensive, affected by the quality of acquired data, and expensive. In recent years, the technology of deep learning has been widely used in the field of seismic exploration. In this paper, deep learning convolutional neural network is introduced, which can build the velocity information of this data directly from seismic data. Attention Unet network distinguishes itself from the traditional network, which can realize the target area by attention according to the observation of demand. Different from the traditional inversion method, the deep learning method is based on the training of big data. In the training phase, the network maps key information from the seismic simulation data into a velocity model. The reconstruction of the data can be completed based on the training. In a large number of experimental data validation results show that the method has achieved better results. The output results can obtain accurate velocity information of underground medium.