Full Waveform Inversion (FWI) is an important method to obtain high-resolution velocity models. However, the seismic data are absence of low-frequency components, the conventional FWI results will have serious cycle skipping problem, which affects the accuracy of the final velocity modeling. For this reason, this paper proposes a local-scale frequency-wavenumber domain phase inversion method, which fully considers the local features of seismic data in the time and offset directions, and utilizes the local-scale decomposition strategy of seismic data and 2D Fourier transform to construct a local-scale frequency-wavenumber domain phase misfit function, to recover the low-wavenumber components of the velocity model and to provide a better initial velocity model for the FWI method. In this paper, we first utilize a 2D sliding window function to extract the local-scale seismic data, and combine it with the 2D Fourier transform to establish the local-scale frequency-wavenumber domain phase information based misfit function. Then, we derived the adjoint-source and gradient operator corresponding to the phase inversion in the local-scale frequency-wavenumber domain. Finally, the test results of the Marmousi model and the igneous-carbonate model show that the local-scale frequency-wavenumber domain phase inversion can provide a better initial velocity model for the conventional FWI method and mitigate the FWI cycle skipping.