1. Number of training images:

The number of training images is determined by the pristineImages datastore. This information can be extracted using:

disp(['Number of pristine images: ', num2str(numel(pristineImages.Files))]);

Based on the code, the output would indicate the total number of pristine images in the specified directory (e.g., "D:\MATLAB\R2024b\Examples\R2024b\iaprtc12\images\39").

2. Input image size:

The input image size is defined in the network using the imageInputLayer:

imageInputLayer([256 256 3], 'Name', 'InputLayer', 'Normalization', 'none')

The input images are resized to 256x256 pixels with 3 color channels (RGB).

b. What are the two types of images used for the training process?

1. Pristine images:

These are the original high-quality images from the pristineImages datastore.

2. Filtered images:

These are the pristine images processed with a bilateral filter, which serve as the ground truth for training. They are stored in the bilatFilteredImages datastore.

c. What are the inputs of the network? What are the desired responses of the network?

1. Inputs of the network:

The noisy image patches extracted from pristineImages.

2. Desired responses of the network:

The corresponding denoised image patches (bilaterally filtered patches) from bilatFilteredImages.

d. What are the layers used in the multiscale CAN from the Deep Learning Toolbox™?

Based on the code, the multiscale CAN includes the following layers:

1. Image input layer:

     Accepts input images of size [256x256x3].

2. Convolutional layers:

     Multiple convolution2dLayer layers with 3x3 filters and padding.

3. ReLU layers:

     Rectified Linear Units (ReLU) are applied after each convolutional layer.

4. Regression layer:

     Outputs the final predicted image values.

e. What is the first layer, middle layer, and last layer of the CAN network in the example?

1. First layer:

imageInputLayer([256 256 3], 'Name', 'InputLayer', 'Normalization', 'none')

2. Middle layer:

For a 10-layer network, the approximate middle layer would be:

convolution2dLayer(3, numberOfFilters, 'Padding', 'same', 'Name', 'Conv3')

reluLayer('Name', 'ReLU3')

3. Last layer:

convolution2dLayer(3, 3, 'Padding', 'same', 'Name', 'ConvFinal')

regressionLayer('Name', 'OutputLayer')

f. What is the training optimizer used to train the network? What are the default values of the hyperparameters of the network?

1. Training optimizer:

The optimizer used is Adam, as specified in the training options:

trainingOptions('adam', ...)

2. Default hyperparameters for Adam optimizer:

Based on the MATLAB Deep Learning Toolbox, the typical default values are:

Learning rate: 0.001 (overridden by your code using 'InitialLearnRate', 1e-3).

Beta1: 0.9

Beta2: 0.999

Epsilon: 1e-8