基于注意力残差网络的结构光包裹相位展开算法

Structure Light Phase Unwrapping Algorithm based on CBAM ResidualNetwork

摘要:

目的：针对传统轮廓条纹测量术包裹相位展开算法中，由于噪声及“孤立区域”等导致的绝对相位误解算的现象，提出了一种基于注意力残差网络的单频绝对相位展开网络。方法：本文解包裹网络的编码器由降采样模块及添加不同填充的残差模块构成，并且将卷积注意力机制模块引入，用来进行特征提取及条纹位置信息学习。解码器由上采样模块及残差模块构成，用来调整网络输出分辨率。结果：本文网络在高斯噪声污染的仿真数据集上的均方根误差比枝切法平均降低了71.4%，解算平均耗时0.4s。本文网络在实测数据集上相对枝切法和最小二乘法解算错误率分别降低了69.9%和43.2%。结论：结果表明该算法能够稳定完成较为复杂曲面和存在“孤立区域”的包裹相位对象的解算。

Abstract:

Objective: In response to the phenomenon of absolute phase miscalculation caused by noise and "isolated regions" in traditional fringe projection-based phase unwrapping algorithms, a single-frequency absolute phase unwrapping algorithm based on CBAM residual network is proposed. Method: The encoder of the unwrapping network in this paper is composed of downsampling modules and residual modules with different padding. Moreover, the convolutional attention mechanism module is introduced to perform feature extraction and learn the position information of the stripes. The decoder consists of upsampling modules and residual modules, which are used to adjust the network output resolution. Result: The root mean square error (RMSE) of the proposed network on a simulated dataset contaminated with Gaussian noise is reduced by an average of 71.4% compared to the branch-cutting method, with an average computation time of 0.4s. On a real-world dataset, the proposed network achieves relative error rate reductions of 69.9% and 43.2% compared to the branch-cutting method and the least squares method, respectively. Conclusion: The results demonstrate that the algorithm can successfully and stably solve the wrapped phase of complex surfaces and objects with "isolated regions".