Electronic speckle fringe denoising using deep learning

Fringe denoising is a critical process for successful interferometry and a long standing issue in automated fringe processing. In this paper, deep learning fringe denoising methods are proposed and applied in electronic speckle pattern interferometry. This paper presents two denoising strategies. The U-Net model is adopted for supervised learning and trained from noisy and clean fringes. In addition, the Noise2Noise method is utilized to train the network from noisy fringe pairs, where the noise of fringe pairs has different distributions. Two datasets are established for training the parameters of two denoising networks. To verify the denoising performance, the fringes with different phase distributions and contrasts are simulated and analyzed. The average mean squared error (MSE) after supervised learning denoising and Noise2Noise denoising method is approximately 0.004 and 0.022, respectively. Different evaluation indicators are used to further evaluate the performance of the two methods. To verify the validity of the measured fringes, the measured fringes are denoised and analyzed using two methods.