DnCNN

DnCNN, short for Denoising Convolutional Neural Network, is a deep learning model introduced in 2017 by Kai Zhang, Wangmeng Zuo, and Lei Zhang for image denoising. It focuses on removing additive white Gaussian noise from images and demonstrates the ability to handle varying noise levels with a single trained model by learning from a mixture of noise intensities.

The architecture centers on a 20-layer convolutional neural network composed of 3x3 filters and 64 feature maps in the hidden layers. The first layer expands the input to 64 channels, the middle layers apply convolution, batch normalization, and ReLU activations, and the final layer outputs a residual image that represents the estimated noise. Training uses a mean squared error loss between the predicted residual and the true noise, so the denoised image is obtained by subtracting the predicted noise from the noisy input. The model supports grayscale and color images by adapting the output to match the input channels.

Training typically uses pairs of clean and synthetically Noised images with Gaussian noise across a range of sigmas, enabling the network to generalize to different noise levels. DnCNN became a widely cited baseline in CNN-based denoising and has influenced subsequent work on blind denoising and related restoration tasks, where networks are designed to estimate and remove noise or unwanted degradation from images. Its residual-learning approach and straightforward architecture have contributed to ongoing developments in deep image restoration.