Prepare Datastore for Image-to-Image Regression

Prepare Data Using Preprocessing Pipeline

This example uses a salt and pepper noise model in which a fraction of input image pixels are set to either 0 or 1 (black and white, respectively). Noisy images act as the network input. Pristine images act as the expected network response. The network learns to detect and remove the salt and pepper noise.

Load Data

Load the pristine images in the digit data set as an imageDatastore. The datastore contains 10,000 synthetic images of digits from 0 to 9. The images are generated by applying random transformations to digit images created with different fonts. Each digit image is 28-by-28 pixels. The datastore contains an equal number of images per category.

digitDatasetPath = fullfile(matlabroot,"toolbox","nnet", ...
    "nndemos","nndatasets","DigitDataset");
imds = imageDatastore(digitDatasetPath, ...
    IncludeSubfolders=true,LabelSource="foldernames");

Specify a large read size to minimize the cost of file I/O.

imds.ReadSize = 500;

Partition Data for Training, Validation, and Testing

Use the shuffle function to shuffle the digit data prior to training.

rng("default")
imds = shuffle(imds);

Use the splitEachLabel function to divide imds into three image datastores containing pristine images for training, validation, and testing.

[imdsTrain,imdsVal,imdsTest] = splitEachLabel(imds,0.95,0.025);

Add Synthetic Noise to Input Images

This example adds synthetic noise to each input image, which will serve as the network input. Define a helper function called addNoise that adds salt and pepper noise to images by using the imnoise (Image Processing Toolbox) function. The addNoise function requires the format of the input data to be a cell array of image data, which matches the format of data returned by the read function of ImageDatastore.

function dataOut = addNoise(data)
    dataOut = data;
    for idx = 1:size(data,1)
       dataOut{idx} = imnoise(data{idx},"salt & pepper");
    end
end

Apply the noise to the input images by using the transform function. The transform function reads data from an underlying datastore and processes the data using the operations defined in the addNoise function. The output of the transform function is a TransformedDatastore.

dsTrainNoisy = transform(imdsTrain,@addNoise);
dsValNoisy = transform(imdsVal,@addNoise);
dsTestNoisy = transform(imdsTest,@addNoise);

Preprocess Data

Use the combine function to combine the noisy images and pristine images into a single datastore that feeds data to trainnet. This combined datastore reads batches of data into a two-column cell array as expected by trainnet. The output of the combine function is a CombinedDatastore.

dsTrain = combine(dsTrainNoisy,imdsTrain);
dsVal = combine(dsValNoisy,imdsVal);
dsTest = combine(dsTestNoisy,imdsTest);

Define a helper function called commonPreprocessing that preprocesses pairs of input and response images. The helper function performs these preprocessing steps.

The helper function requires the format of the input data to be a two-column cell array of image data, which matches the format of data returned by the read function of CombinedDatastore.

function dataOut = commonPreprocessing(data)
    dataOut = cell(size(data));
    for col = 1:size(data,2)
        for idx = 1:size(data,1)
            temp = single(data{idx,col});
            temp = imresize(temp,[32,32]);
            temp = rescale(temp);
            dataOut{idx,col} = temp;
        end
    end
end

Apply the preprocessing to the input and response data by using the transform function. Preprocess the training, validation, and test data sets.

dsTrain = transform(dsTrain,@commonPreprocessing);
dsVal = transform(dsVal,@commonPreprocessing);
dsTest = transform(dsTest,@commonPreprocessing);

Augment Training Data

Augmentation reduces overfitting and adds robustness to the presence of rotations in the trained network. Randomized augmentation is not needed for the validation or test data sets.

Define a helper function called augmentImages that adds randomized 90 degree rotations to the data by using the rot90 function. Identical rotations are applied to the network input and corresponding expected responses. The function requires the format of the input data to be a two-column cell array of image data, which matches the format of data returned by the read function of CombinedDatastore.

function dataOut = augmentImages(data)
    dataOut = cell(size(data));
    for idx = 1:size(data,1)
        rot90Val = randi(4,1,1)-1;
        dataOut(idx,:) = {rot90(data{idx,1},rot90Val), ...
            rot90(data{idx,2},rot90Val)};
    end
end

Apply the randomized augmentation to the training data by using the transform function.

dsTrain = transform(dsTrain,@augmentImages);

Preview Preprocessed Data

Since there are several preprocessing operations necessary to prepare the training data, preview the preprocessed data to confirm it looks correct prior to training. Use the preview function to preview the data.

Visualize examples of paired noisy and pristine images using the montage (Image Processing Toolbox) function. The training data looks correct. Salt and pepper noise appears in the input images in the left column. Other than the addition of noise, the input image and response image are the same. Randomized 90 degree rotation is applied to both input and response images in the same way.

exampleData = preview(dsTrain);
inputs = exampleData(:,1);
responses = exampleData(:,2);
minibatch = cat(2,inputs,responses);
montage(minibatch',Size=[8 2])
title("Inputs (Left) and Responses (Right)")

Define Convolutional Autoencoder Network

Convolutional autoencoders are a common architecture for denoising images. Convolutional autoencoders consist of two stages: an encoder and a decoder. The encoder compresses the original input image into a latent representation that is smaller in width and height, but deeper in the sense that there are many feature maps per spatial location than the original input image. The compressed latent representation loses some amount of spatial resolution in its ability to recover high frequency features in the original image, but it also learns to not include noisy artifacts in the encoding of the original image. The decoder repeatedly upsamples the encoded signal to move it back to its original width, height, and number of channels. Since the encoder removes noise, the decoded final image has fewer noise artifacts.

This example defines the convolutional autoencoder network using layers from Deep Learning Toolbox™, including:

Create the image input layer. To simplify the padding concerns related to downsampling and upsampling by factors of two, choose a 32-by-32 input size because 32 is cleanly divisible by 2, 4, and 8.

imageLayer = imageInputLayer([32,32,1]);

Create the encoding layers. Downsampling in the encoder is achieved by max pooling with a pool size of 2 and a stride of 2.

encodingLayers = [ ...
    convolution2dLayer(3,8,Padding="same"), ...
    reluLayer, ...
    maxPooling2dLayer(2,Padding="same",Stride=2), ...
    convolution2dLayer(3,16,Padding="same"), ...
    reluLayer, ...
    maxPooling2dLayer(2,Padding="same",Stride=2), ...
    convolution2dLayer(3,32,Padding="same"), ...
    reluLayer, ...
    maxPooling2dLayer(2,Padding="same",Stride=2)];

Create the decoding layers. The decoder upsamples the encoded signal using a transposed convolution layer with a stride of 2, which upsamples by a factor of 2. The network uses a clippedReluLayer as the final activation layer to force outputs to be in the range [0, 1].

decodingLayers = [ ...
    transposedConv2dLayer(2,32,Stride=2), ...
    reluLayer, ...
    transposedConv2dLayer(2,16,Stride=2), ...
    reluLayer, ...
    transposedConv2dLayer(2,8,Stride=2), ...
    reluLayer, ...
    convolution2dLayer(1,1,Padding="same"), ...
    clippedReluLayer(1.0)];    

Concatenate the image input layer, the encoding layers, and the decoding layers to form the convolutional autoencoder network architecture.

layers = [imageLayer,encodingLayers,decodingLayers];

Define Training Options

Train the network using Adam optimization. Specify the hyperparameter settings by using the trainingOptions function. Train for 50 epochs.

options = trainingOptions("adam", ...
    MaxEpochs=50, ...
    MiniBatchSize=imds.ReadSize, ...
    ValidationData=dsVal, ...
    ValidationPatience=5, ...
    Plots="training-progress", ...
    OutputNetwork="best-validation", ...
    Verbose=false);

Train the Network

Train the neural network using the trainnet function. For regression, use mean squared error loss. By default, the trainnet function uses a GPU if one is available. Training on a GPU requires a Parallel Computing Toolbox™ license and a supported GPU device. For information on supported devices, see GPU Computing Requirements (Parallel Computing Toolbox). Otherwise, the trainnet function uses the CPU. To specify the execution environment, use the ExecutionEnvironment training option. Training takes about 15 minutes on an NVIDIA Titan XP.

net = trainnet(dsTrain,layers,"mse",options);

modelDateTime = string(datetime("now",Format="yyyy-MM-dd-HH-mm-ss"));
save("trainedImageToImageRegressionNet-"+modelDateTime+".mat","net");    

Evaluate the Performance of the Denoising Network

Obtain output images from the test set by using the minibatchpredict function.

ypred = minibatchpredict(net,dsTest);

Obtain pairs of noisy and pristine images from the test set using the preview function.

testBatch = preview(dsTest);

Visualize a sample input image and the associated predicted output from the network to get a sense of how well denoising is working. As expected, the output image from the network has removed most of the noise artifacts from the input image. The denoised image is slightly blurry as a result of the encoding and decoding process.

idx = 1;
y = ypred(:,:,:,idx);
x = testBatch{idx,1};
ref = testBatch{idx,2};
montage({x,y})

Assess the performance of the network by analyzing the peak signal-to-noise ratio (PSNR). The PSNR of the output image is higher than the noisy input image, as expected.

psnrNoisy = psnr(x,ref)

psnrNoisy = single
    20.8214

psnrDenoised = psnr(y,ref)

psnrDenoised = single
    21.5986