Phase unwrapping plays an important role in many applications, but it is always affected by noise, especially speckle noise. In order to remove the effect of speckle noise on the wrapping phase diagram, the actual phase value was recovered from it. The two-step method was used to make theoretical analysis and experimental demonstration, and the absolute phase was recovered from the phase covered by speckle noise in different degrees. In the first step, based on the swin-UNet-denoise network, the normalized layer in swin block was set back, and the attention value was calculated by cosine similarity. Then, the relative displacement offset was replaced by logarithmic position offset, and the deconvolution layer was fused in the upsampling module to improve the denoising ability of the network. In the second step, the denoising result was unwrapped by least square method, and then the absolute phase was obtained by median filtering. The results show that the structural similarity is 99.77%, the peak signal-to-noise ratio is 39.98, the root-mean-square error is 0.4864, and the average absolute error is 0.4302, respectively. In addition, all the networks are only trained and verified on 300 simulation data sets, which proves that the research can provide a reference for faster, more efficient and accurate phase unwrapping with speckle noise even under the condition of small samples.