The proposed invention is a system and method for training a Convolutional Neural Network (CNN) to predict a tuning parameter to be used in an existing image denoising method (called BM3D) to obtain best possible denoising results on images obtained by digital cameras in low-light conditions. The performance of the BM3D denoising algorithm varies with this tuning parameter.
In this work we present a method to predict the best parameter value for each image patch and we observe that using this prediction we obtain better results than using a fixed parameter value for all images.
There are many image denoising methods available today. However, they are trained and tested on artificial noise. According to our observations, when it comes to images corrupted by real low light noise the BM3D method works best. Our work takes the BM3D method and enhances it by predicting what its tuning parameter should be for each image patch being denoised.
This technology could be directly sold to consumers in the form of an app or embedded in a mobile phone or digital camera.
