Two-dimensional phase unwrapping algorithms are widely used in optical metrology-related fields. However，complex environments such as high noise and phase discontinuity in practical application scenarios often lead to the failure of traditional phase unwrapping. In this paper，a method based on deep convolutional neural network（DCNN）is proposed for phase unwrapping，which considers phase unwrapping as a multi-pixel classification problem and introduces an improved ResUNet segmentation network to recognize the categories，and after the segmentation is completed，the unwrapped phase map is combined with the segmentation result to generate the unwrapped phase. Once the segmentation is completed，the unwrapped phase can be generated by combining the parcel phase map and the segmentation result. In this paper，we compare with the existing methods on simulation datasets for the noise and discontinuity cases，respectively，and the phase unwrapping RMSE is only 0.006 2 for the wrapped phase map with -2 dB noise level，and for the phase discontinuity case，the RMSEm and RMSEsd are 0.001 7 and 0.017 8，which are much lower than ResUNet and several other methods.