Denoising method based on an improved convolutional neural network for laser speckle pattern interferometry phase maps with high noise levels

Authors:

Abstract
Denoising of speckle pattern interference phase maps is a complex but critical step in laser speckle pattern interferometry. This study proposes an algorithm to denoise high-noise phase fringe patterns on the basis of an improved convolutional neural network. The fundamental structure of the algorithm is a U-Net configuration with residual connections. These connections help solve the problem of gradient vanishing during the training of deep neural networks. During algorithm design, the model is trained by combining the simulated noise fringe patterns and the real noise fringe patterns to construct a training set with a certain proportion, which improves the generalization ability of the denoising algorithm. Different from the single-indicator loss function in the traditional neural network denoising model, the combined loss function proposed in this study merges the structured similarity indexing measure and mean square error as evaluation indices in the training process of the algorithm model. The combined loss function can accurately evaluate the similarity between the denoised image and the original image and reduce the errors that may result from a single index. The algorithm, with its excellent predictive analysis and feature extraction capabilities, efficiently removes noise from phase-stripe images and preserves the edge and detailed information of the stripes. Test results show that the proposed algorithm can achieve good denoising effects and retain the fringe detail information of interferometric fringe patterns when dealing with speckle pattern interference phase maps with sparse fringe density and low noise and those with dense and noisy fringes. This noise reduction effect is crucial for improving the quality of interferometric phase maps and lays a solid foundation for subsequent interferometric phase map processing and analysis, such as phase unwrapping.
Funding Information
  • Planning Issues of Philosophy and Social Sciences (23LMJX16YB)
  • Zhejiang Provincial Postdoctoral Research Project Selection and Optimization Funding (2024-00004)
  • Ningbo Public Welfare Research Program (2023S090)