Compared with traditional iterative methods, deep learning phase reconstruction has lower bit error rate and higher data transfer rate. We found the efficiency of training mainly was from the edges of the phase patterns due to their stronger intensity changes between adjacent phase distribution. According to this characteristic, we proposed a method to only record and use the high frequency component of the phase patterns and to do the deep learning training. This method can improve the storage density due to reducing the material consumption.

The phase retrieval method based on deep learning can be used to solve the iterative problem in holographic data storage. The key of the deep learning method is to build the relationship between the phase data pages and the corresponding near-field diffraction intensity patterns. However, to build the correct relationship, thousands of samples of the training dataset are usually required. In this paper, according to the coding characteristics of phase data pages, we proposed an image segmentation method to greatly reduce the number of original training dataset. The innovation proposed by this new method lies in the special segmentation of the original samples to expand the number of samples.