Image decomposition with discrete wavelet transform to design a denoising neural network

Reducing noise in digital images is one of the most common tasks in image processing. At the moment, noise reduction approaches based on the applying of convolutional neural networks are widely used. In this case, as a rule, model training is based on minimizing the error function between the result of the network operation and the expected reference image and, additionally, various representations of the two-dimensional image signal and their properties are not used to optimize the training of noise reduction network architectures. The paper proposes an approach to training neural networks to suppress noise. The described approach is based on the usage of the N-fold fast Haar wavelet transform. This representation of a discrete image signal allows one to discard the classical architecture of the autoencoder and to use only its part that encodes the signal which leads to a significant reduction in model parameters and speeds up the network.