How to evaluate the image generated with CycleGAN?

A network is trained using unsupervised data from both a 'good' image dataset and a 'noisy' image dataset with GAN. Subsequently, the trained network reconstructs the noisy image, resulting in a 'generated image' or 'denoised image'. At this point, we have both the 'noisy image' and the 'generated image,' with no 'real image' or 'target image' available.

Traditional image evaluation functions like "ssim(generated images, target image)" (and many other image evaluation functions) may not be suitable in this scenario for comparing the 'noisy image' and the 'generated image.' Also, metrics such as 'Frechet Inception Distance (FID)' and 'Inception Score (IS)' are commonly used to assess the distribution to evaluate the GAN model.

Given this context, what would be the most appropriate evaluation methods to measure the quality of images generated by the trained GAN network, with only the 'noisy-image' and the 'generated-image' (and the trained network)?

Your insights and wisdom on this matter are highly appreciated.

Thank you

Answer 1

[Ayush Modi]

Hi John,

You can consider to assess the performance of the GAN model using "No Reference Image Quality" Metrics. Most commonly used "No Reference Image Quality Metrics" are:

• BRISQUE - Blind/Referenceless image spatial quality evaluator
• NIQE - Naturalness image quality evaluator
• PIQE - Perception-based image quality evaluator

Please refer to the following MathWorks documentation for more information on the "No Reference Image Quality" metrics:

https://www.mathworks.com/help/images/compare-no-reference-image-quality-metrics.html

Note - Each metric has different strengths depending on the images in the data set. To select the best metric for your data, you can compare the performance of the three metrics on sample image data.

Answer 2

[John]

Thank you, Ayush, for your valuable input.

It would be beneficial to have an evaluation criterion that considers both images. In the context of medical images, the expectation is that all features are preserved, and no new features are introduced after any processing, such as denoising.

Best regards.

[John]

Ayush:

I found the fitbrisque and fitniqe to build the customer evaluation model. However, I cannot find the fitpiqe to create the Perception-based image quality evaluator. Do you know whether there is one?