Digital holographic imaging is a key technology to obtain wave front information of 3D objects, and obtaining high-quality holograms is the first condition. However, due to the constraints of the imaging sensor and the influence of the experimental environment, the obtained digital hologram is noisy and has low resolution, which affects the SNR in the holographic reconstruction. To overcome this constraint, a method based on deep learning is used to improve the hologram quality and signal-to-noise ratio. The results show that the algorithm can be used in the acquisition of multi-scale hologram, and obtain high quality hologram reconstruction effect is better, to reduce speckle noise, and compare the three kinds of loss function in the performance of network training.