Current data-driven deep learning (DL) methods typically reconstruct subsurface velocity models directly from pre-stack seismic records. However, these purely data-driven methods are often less robust and produce results that are less physically interpretative. Here, the authors propose a new method that uses migration images as input, combined with convolutional neural networks to construct high-resolution velocity models. Compared to directly using pre-stack seismic records as input, the nonlinearity between migration images and velocity models is signiffcantly reduced. Additionally, the advantage of using migration images lies in its ability to more comprehensively capture the reffective properties of the subsurface medium, including amplitude and phase information, thereby to provide richer physical information in guiding the reconstruction of the velocity model. This approach not only improves the accuracy and resolution of the reconstructed velocity models, but also enhances the physical interpretability and robustness. Numerical experiments on synthetic data show that the proposed method has superior reconstruction performance and strong generalization capability when dealing with complex geological structures, and shows great potential in providing efffcient solutions for the task of reconstructing high-wavenumber components.