Residual Curvature Analysis (RCA) based on common image point gathers(offset domain) can significantly improve the accuracy of migration velocity modeling and Pre-Stack Time Migration (PSTM). By picking up the points with the maximum energy in the spectrum data, one can achieve a new time domain migration velocity. However, the traditional methods for instance: migration velocity modeling based on manual picking of γ spectrum, or migration velocity analyzing based on Human-Computer Interaction etc. Which are very labor-intensive and time-consuming. To help address this concern, we proposed a kind of migration velocity modeling method based on deep learning. First, A series of random γ values are applied to the target velocity model, and after performing PSTM, then common image point gathers are obtained. Second, based on common image point gathers, we can calculate corresponded γ spectrum through γ scanning, which can be taken as the sample data-set and input γ can be used as label data. In terms of constructing a γ value prediction network, we implement a neural network based on the architectures of U-Net3+, and we chose Mish as the activation function, and we chose Log_Cosh as the misfit function. We can train the network using data-sets above, and corrected migration velocity model can be predicted though inferring input γ spectrum. Method test based on Marmousi model has verified the correctness of the method proposed in this paper. And a large amount of field data processing results show that the method proposed in this paper can obtain more accuracy velocity model than manual picking, while its efficiency is dozens of times higher than manual picking, which turns out our method has great potential in industrial application.Although certain achievements have been made in this paper, continuous efforts are still required in terms of the generalization of the neural network model, the pre-processing of CRP gather data, and the completeness of the data-set.