Continuous Phase Denoising via Deep Learning Based on Perlin Noise Similarity in Digital Holographic Microscopy

Deep learning (DL), as a powerful tool for image processing by learning from data, brings significant advantages for solving the noise in phase images of digital holography. However, due to the inaccurate representation of the features of the actual digital holographic hybrid phase noise (DHHPN), most DL-based denoising strategies that use a Gaussian noise model to generate simulated datasets for training have unsatisfactory performance. Here, to explore the characteristics of actual DHHPN, we evaluate the correlations between the DHHPN obtained from different DH/DHM experiments and five types of noise (Uniform, Normal, Pink, Brown and Perlin) in terms of maximum information coefficient and Pearson correlation coefficient. For the first time, to the best of our knowledge, we have revealed an extremely high similarity between DHHPN and Perlin noise. Based on this discovery, a continuous phase denoising method via deep learning based on Perlin noise similarity is proposed. Without needing to collect and label experimental training data at high cost, a dataset consisting only of computer-generated clean sample images and Perlin noise images can be easily obtained. This simulated dataset is then used to train our designed convolutional neural network. Simulation and experimental results show that the denoising performance of the proposed method far exceeds the other two classical methods, and the standard deviation of the measurement results is reduced by an order of magnitude, reaching the sub-nanometer level. The proposed method has significant application in the fields of digital holographic precision measurements.