Phase unwrapping occupies an important position in digital holographic microscopy imaging technology. It is an indispensable key step for obtaining phase information. The current traditional phase unwrapping algorithm has entered the platform stage. Deep learning for phase unwrapping research methods were used and better experimental results were obtained. This study proposes an improved U-Net phase unwrapping method, adding the channel attention module after the residual block, and uses deep separable convolution to replace part of the traditional convolution. This study uses a large number of analog phase maps generated by random matrix as data training sets to achieve the purpose of phase unwrapping, and has solved the problem that optical path is inconvenient to obtain a large number of data sets. The proposed method is used to unpack different experimental holograms, and compared with the results of other unpacking algorithms, the experimental results show that the edge of the proposed method is smoother and the background is flatter. The structural similarity index of the experimental test set and the Discrete Cosine Transform unwrapping results increased from an average of 0.932 to 0.973, and the peak signal-to-noise ratio increased from an average of 21.60 to 29.18.