U-Net based neural network for fringe pattern denoising

•In this paper, we present a residual dense convolutional neural network fringe pattern image denoising.•We propose a denoising block that uses densely connected grouped convolutional blocks.•The proposed model adopts the basic U-net auto-encoder and the use of specialized denoising blocks, but with lower computational cost and better performance.•Our proposal is a lightweight residual dense U-net (LRDUNet) that combines densely connected grouped convolutional blocks.•The model allows us the reuse of the feature maps within convolutional blocks local residual learning to address the vanishing gradient problem, and also includes a global residual learning to estimate the noise of the image instead of the denoised image directly. Fringe patterns from different optical measurement systems are widely used in scientific and engineering applications. However, fringe patterns are often corrupted by speckle noise, which is necessary to be removed to accurately recover the information encoded in the phase of the fringe pattern. In this paper we propose a lightweight residual dense neural network based on the U-net neural network model (LRDUNet) for fringe pattern denoising. The encoding and decoding layers of the LRDUNet consist of grouped densely connected convolutional layers for the sake of reusing the feature maps and reducing the number of trainable parameters. Additionally, local residual learning is used to avoid the vanishing gradient problem and speed up the learning process. We compare the proposed method versus state-of-the-art methods and present a study of parameters where we demonstrate that computationally simpler versions of the proposed model are still quite competitive. Experiments on simulated and real fringe patterns show that the proposed method outperforms state-of-the-art methods by restoring the main features of the fringe patterns, achieving an average of 41 dB of PSNR on simulated images.