研究目的
To reduce the noise of fringe patterns using deep learning, enabling faster and higher-quality denoising compared to existing methods, with potential applications in other optical denoising problems.
研究成果
The proposed deep learning-based denoising method effectively reduces noise in fringe patterns with faster computation speed and better performance at boundaries compared to existing algorithms like WFF. It demonstrates high-quality results on both simulated and real data, offering a stable solution without parameter adjustments and potential for broader optical applications.
研究不足
The method relies on simulated training data, which may not fully capture all real-world noise characteristics. The DCNN model is trained for specific noise levels and fringe pattern sizes, potentially limiting generalization to other conditions. Computational resources required for training (e.g., GPU) may be a constraint.
1：Experimental Design and Method Selection:
A deep convolutional neural network (DCNN) model is designed with 20 convolutional layers, using ReLU activation functions and Adam optimization algorithm for training. The method involves simulating training data to avoid experimental acquisition.
2：Sample Selection and Data Sources:
Training samples are generated using simulations based on the fringe pattern equation with Gaussian noise added. Zernike polynomials are used to model wavefronts. Real fringe patterns are collected from interferometers (Zygo and 4D) for testing.
3：List of Experimental Equipment and Materials:
A PC with Intel Core i7-7820X CPU and NVIDIA GeForce GTX 1080 GPU is used for computation. Software includes Python and PyTorch framework.
4：Experimental Procedures and Operational Workflow:
The DCNN is trained on 80,000 simulated noisy and noiseless fringe pattern pairs. Training involves multiple epochs with decreasing learning rates. Testing is performed on both simulated and real fringe patterns, comparing results with the windowed Fourier filter (WFF) algorithm.
5：Data Analysis Methods:
Performance is evaluated using visual observation, mean square error (MSE) calculations, and error curve analysis to compare denoising results with ground truths.
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
