EValuate the generated images of GANS using FID and SSIM (MATLAB:2023a)

I have generated an images using CGANS ( conditional Gans) as in the matlab official example:

https://se.mathworks.com/help/deeplearning/ug/train-conditional-generative-adversarial-network.html?searchHighlight=cgans&s_tid=srchtitle_support_results_1_cgans

Howerver, I need to write a code to evaluate the smilarity between the orignial images and generated images using FID (Frechet Inception Distance (FID) or Image Quality Measures (SSIM)

[Fatima Bibi]

i also want this question answer i want to load cgan pretrained model that matches model images with original images

my original images have one folder that contain 5 classes.i want it matches randomly 10 images per class for FID

Answer 1

[Ayush Aniket]

Hi MAHMOUD EID,

For calculating SSIM for the generated images you can use the MATLAB inbuilt ‘ssim’ function as follows:

% Load and preprocess the original and generated images
original_images = imread('path/to/original/image.jpg');
generated_images = imread('path/to/generated/image.jpg');

% Calculate Structural Similarity Index Measure (SSIM)
ssim_score = ssim(original_images, generated_images);

You can read more about the function in the documentation page: https://in.mathworks.com/help/images/ref/ssim.html

For calculating FID, you can use pre-trained Inception-v3 model from the Deep Learning Toolbox as follows:

net = inceptionv3;

% Calculate Frechet Inception Distance (FID)
original_features = activations(net, original_image, 'avg_pool');
generated_features = activations(net, generated_image, 'avg_pool');
fid_score = wassersteinDistance(original_features, generated_features);

You can read about the model here : https://in.mathworks.com/help/deeplearning/ref/inceptionv3.html

You may need the check for the size compatibility of your images and that required as input to the Inception-v3 model.

Hope this helps!

[MAHMOUD EID]

Hi, @Ayush Aniket

Thanks for the frutiful explination

for FID method , there is no wassersteinDistance function in matlab.

So can you upload the code of this function?

[Shreeya]

Hi @MAHMOUD EID

Find the implementation of the wassersteinDistance below

Wasserstein distance - File Exchange - MATLAB Central (mathworks.com)

[John]

Is the 'original image' before the 'net'? For example, is it the 'noised image' before denoising? If that is the case, then "ssim(original_images, generated_images);" will not be much meaningful. I am not sure how about the Frechet Inception Distance (FID).

What kind of appropriate quantitative evaluation for, for example, unsupervised Cycle GAN generated denoised images? There are two group of images here: noisy-image and denoised-image; the training net is available.

Thank you.