A Document Image Moiré Pattern Reduction Method Based on Convolutional Neural Networks

In recent years, the removal of Moiré patterns from images has been a challenging problem in computer vision, especially for document images. Many deep learning-based image denoising methods have been proposed to address the problem of Moiré pattern removal, but these methods still have issues with recovering textural details and globally removing Moiré patterns. This paper proposes a coarse-to-fine two-stage convolutional neural network to address the problems of incomplete Moiré pattern removal and loss of textural details. In the first stage, the model estimates the positions of the Moiré patterns and performs preliminary removal. In the second stage, the focus is on optimizing the preliminary removal results, which avoids the problems of incomplete removal and loss of textural details. In particular, we replace the loss function in the second stage with a focus on improving image quality, and add a Non-Local module to increase the receptive field, paying more attention to the information around the Moiré patterns. Finally, experimental results on the public dataset of Baidu PaddlePaddle demonstrate that our model outperforms other convolutional neural network models.