研究目的
To propose a novel framework for unwrapping the phase using deep fully convolutional neural network termed as PhaseNet, reformulating the problem definition of directly obtaining continuous original phase as obtaining the wrap-count at each pixel by semantic segmentation.
研究成果
The proposed PhaseNet framework effectively applies Deep Convolutional Neural Network for phase unwrapping, demonstrating robustness to noise and computational efficiency. It reformulates the problem as semantic segmentation, leveraging the relationship between absolute phase and wrap-count for training data generation. The framework outperforms traditional methods in terms of noise robustness and computational time, paving the way for new deep learning based phase unwrapping methods.
研究不足
The PhaseNet is trained with randomly varying shapes generated by combining Gaussians, which may limit its adaptability to very specific or irregular patterns not represented in the training data.
1：Experimental Design and Method Selection:
The proposed framework uses a deep fully convolutional neural network (PhaseNet) for phase unwrapping, reformulating it as a semantic segmentation problem.
2：Sample Selection and Data Sources:
Training data is generated by performing arithmetic operations on Gaussian functions with randomly varying mean and variance values, adding Gaussian noise for practicality.
3：List of Experimental Equipment and Materials:
NVIDIA GTX 1080-Ti GPU with 11 GB memory for training.
4：Experimental Procedures and Operational Workflow:
The PhaseNet architecture consists of an encoder network, a corresponding decoder network followed by a pixel-wise classification layer. Training involves cross-entropy loss and Adam optimizer.
5：Data Analysis Methods:
Performance is compared with quality-guided phase unwrapping algorithm and MATLAB’s unwrap function for varying noise levels using Mean Square Error (MSE) and processing time.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
