To address the problem of a single deep-learning model being unable to reconstruct the wavefront of digital holograms with multiple scales, an improved network structure based on the U-Net model is proposed to simulate the digital holographic imaging process and generate holographic images with different scales as data sets. Digital holograms with different scales are used in different parts of the training network, and a depth learning model is obtained, which can reconstruct the wavefront information of digital holograms with three different scales. The experimental results show that the proposed network structure can reconstruct digital holograms with various scales and obtain accurate wavefront information of digital holograms. The research content solves the problem of using a single deep-learning model to deal with digital holograms with varying scales.