Vol. 54, Issue 4, pp. 523-539 (2024)

Abstract

In order to solve the problem of unwrapping the phase in digital holographic reconstruction, a new convolutional architecture is proposed. The proposed method takes the U-net network as a framework, and incorporates the lightweight deep learning network of Mobilenetv1 in the encoding part to reduce the model complexity, the number of parameters, and the cost of computation; and proposes a complex dual-channel convolutional block in the decoding part instead of the 3 × 3 convolution in the original U-net network, which fully incorporates the features in the decoding process. (abbreviated as DC-UMnet network) Meanwhile, the loss value is calculated using the SmoothL1Loss function, and the activation function uses ReLU6. Finally, the simulated dataset containing noise is used for training; and the experimentally obtained wrapped phase maps is used for verifying. The simulation results show that under different degrees of noise conditions, compared with the DCT method and the deep learning phase unwrapping network, the structural similarity index values of the DC-UMnet network are improved by 0.416 and 0.064; and the normalized root-mean-square errors are reduced respectively by 13.2% and 5.8%. Through the actual measurement data, the proposed network model of the feasibility and good noise reduction ability are verified, which can realize digital holographic phase unwrapping in a simple, fast and efficient way.