Zhang Junkang, Li Qingguang
Opt Express. 2022 Mar 28;30(7):10470-10490. doi: 10.1364/OE.444875.
In this paper, we first propose a quantitative indicator to measure the amount of prior information contained in the wrapped phase map. Then, Edge-Enhanced Self-Attention Network is proposed for two-dimensional phase unwrapping. EESANet adopts a symmetrical en-decoder architecture and uses self-designed Serried Residual Blocks as its basic block. We add Atrous Spatial Pyramid Pooling and Positional Self-Attention to the network to obtain the long-distance dependency in phase unwrapping, and we further propose Edge-Enhanced Block to enhance the effective edge features of the wrapped phase map. In addition, weighted cross-entropy loss function is employed to overcome the category imbalance problem. Experiments show that our method has higher precision, stronger robustness and better generalization than the state-of-the-art.
在本文中，我们首先提出一种定量指标来衡量包裹相位图中包含的先验信息量。然后，提出了边缘增强自注意力网络用于二维相位解缠。EESANet采用对称的编码器-解码器架构，并使用自行设计的密集残差块作为其基本模块。我们在网络中添加空洞空间金字塔池化和位置自注意力以获得相位解缠中的长距离依赖性，并且进一步提出边缘增强块来增强包裹相位图的有效边缘特征。此外，采用加权交叉熵损失函数来克服类别不平衡问题。实验表明，我们的方法比现有技术具有更高的精度、更强的鲁棒性和更好的泛化能力。
