Encoder-Decoder Model for Grayscale to RGB Conversion

This script demonstrates creating and training an encoder-decoder model using random overlapping image patches.

This script demonstrates how to create random overlapping patches of an image to train an encoder-decoder model that converts grayscale images to RGB images.

Some steps:

• reading an example image (`peppers.png`)
• call the function `image2patch_randperm_overlap` wiht the parameter

clc
clear
close all force
X = imread('peppers.png') ;
figure, imagesc(X), title('Example')

imageOut = mat2gray( X ) ;
imageIn  = rgb2gray( imageOut ) ;

deltaX = 32 ;
deltaY = 32 ;
overlap = 8 ;

per = .8 ; % percent to train
seed = 1 ; % to control the random numbers
make_plot = 1 ;

[patchs] = image2patch_randperm_overlap( ...
    imageIn , imageOut , deltaX , deltaY , overlap , seed , per , make_plot) ;

• to visualizate the Train and Test data:

• Encoder-Decoder Model: created using the Deep Neural Network designer

net = trainNetwork(patchs.train.In,patchs.train.Out,lgraph,options);

• Computing the SSIM value (in average for the patchs) to get an idea of how well is the adjusted the model:

% SSIM_Training_Data = 0.8577
% SSIM_Testing_Data = 0.7857

• Testing on the whole image (including the training patches):

• Testing on a different Image (`cameraman.tif`):

Key Functions and Parameters

- **image2patch_randperm_overlap**: Generates random overlapping patches from the input images.

- **layer_Encoder_Decoder**: Script that defines the encoder-decoder network architecture.

- **trainingOptions**: Configures the training options for the network.

- **trainNetwork**: Trains the neural network.

- **mean_ssim**: Calculates the mean SSIM between predicted and actual images (for 4D matrices).

- **image2patch_overlap**: Converts an image into overlapping patches.

- **patch_overlap2image**: Converts patches back into a full image.

Conclusion

This script effectively demonstrates the process of creating and training an encoder-decoder model using random image patches and testing the model on both the training image and a new test image. The use of overlapping patches helps in reconstructing the full image during the prediction stage. This approach exemplifies how a deep learning model can be trained for the challenging task of converting grayscale images to RGB using a single image without repeating patches.

Author: Michael Alvarez

Email: [e-mail]