Main Content

Train and Apply Denoising Neural Networks

Image Processing Toolbox™ and Deep Learning Toolbox™ provide many options to remove noise from images. The simplest and fastest solution is to use the built-in pretrained denoising neural network, called DnCNN. However, the pretrained network does not offer much flexibility in the type of noise recognized. For more flexibility, train your own network using predefined layers, or train a fully custom denoising neural network.

Remove Gaussian Noise Using Pretrained Network

You can use the built-in pretrained DnCNN network to remove Gaussian noise without the challenges of training a network. Removing noise with the pretrained network has these limitations:

  • Noise removal works only with 2-D single-channel images. If you have multiple color channels, or if you are working with 3-D images, remove noise by treating each channel or plane separately. For an example, see Remove Noise from Color Image Using Pretrained Neural Network.

  • The network recognizes only Gaussian noise, with a limited range of standard deviation.

To load the pretrained DnCNN network, use the denoisingNetwork function. Then, pass the DnCNN network and a noisy 2-D single-channel image to denoiseImage. The image shows the workflow to denoise an image using the pretrained DnCNN network.

The denoiseImage function removes noise from a grayscale image using a pretrained denoising network.

Train a Denoising Network Using Built-In Layers

You can train a network to detect a larger range of Gaussian noise standard deviations from grayscale images, starting with built-in layers provided by Image Processing Toolbox. To train a denoising network using predefined layers, follow these steps. The diagram shows the training workflow in the dark gray box.

  • Create an ImageDatastore object that stores pristine images.

  • Create a denoisingImageDatastore object that generates noisy training data from the pristine images. To specify the range of the Gaussian noise standard deviations, set the GaussianNoiseLevel property. You must use the default value of PatchSize (50) and ChannelFormat ('grayscale') so that the size of the training data matches the input size of the network.

  • Get the predefined denoising layers using the dnCNNLayers function.

  • Define training options using the trainingOptions (Deep Learning Toolbox) function.

  • Train the network, specifying the denoising image datastore as the data source for trainNetwork (Deep Learning Toolbox). For each iteration of training, the denoising image datastore generates one mini-batch of training data by randomly cropping pristine images from the ImageDatastore, then adding randomly generated zero-mean Gaussian white noise to each image patch. The standard deviation of the added noise is unique for each image patch, and has a value within the range specified by the GaussianNoiseLevel property of the denoising image datastore.

After you have trained the network, pass the network and a noisy grayscale image to denoiseImage. The diagram shows the denoising workflow in the light gray box.

The denoiseImage function removes noise from a grayscale image using a denoising network that you train.

Train Fully Customized Denoising Neural Network

To train a denoising neural network with maximum flexibility, you can use a custom datastore to generate training data or define your own network architecture. For example, you can:

  • Train a network that detects a larger variety of noise, such as non-Gaussian noise distributions, in single-channel images. You can define the network architecture by using the layers returned by the dnCNNLayers function. To generate training images compatible with this network, use the transform and combine functions to batches of noisy images and the corresponding noise signal. For more information, see Preprocess Images for Deep Learning (Deep Learning Toolbox).

    After you train a denoising network using the DnCNN network architecture, you can use the denoiseImage function to remove image noise.

    Tip

    The DnCNN network can also detect high-frequency image artifacts caused by other types of distortion. For example, you can train the DnCNN network to increase image resolution or remove JPEG compression artifacts. The JPEG Image Deblocking Using Deep Learning example shows how to train a DnCNN network to remove JPEG compression artifacts

  • Train a network that detects a range of Gaussian noise distributions for color images. To generate training images for this network, you can use a denoisingImageDatastore and set the ChannelFormat property to 'rgb'. You must define a custom convolutional neural network architecture that supports RGB input images.

    After you train a denoising network using a custom network architecture, you can use the activations (Deep Learning Toolbox) function to isolate the noise or high-frequency artifacts in a distorted image. Then, subtract the noise from the distorted image to obtain a denoised image.

See Also

| | | (Deep Learning Toolbox) | (Deep Learning Toolbox) | | (Deep Learning Toolbox) | |

Related Examples

More About