Denoising image data overwhelmed by noise is a challenging task. How do you handle it with machine learning if you don’t have clean “target” images to learn a mapping from noisy to clean data? Our new publication “Denoising image-based experimental data without clean targets based on deep autoencoders” examines this research question.
We propose a denoising blind autoencoder, which does not need any clean images for training and can be used in test cases with poor signal/noise ratio (for instance fast Pressure Sensitive Paints). The method is expected to extend our capability of measuring wall quantities, which is fundamental in flow investigation and control.
Read the open-access paper here.
The data and code are also freely availiable.
This work is a collaboration with Feng Gu, Yingzheng Liu, Zhaomin Cao, and Di Peng of Shanghai Jiao Tong University.
