DCT2net software: Trained shallow CNN-based DCT denoiser
Description DCT2Net method. a, Architecture of DCT2net for a patch size p=5. b, Original DCT basis (left) vs DCT2net learned basis (right). c, Denoising results (with PSNR values) of Castle image fromBSD68 dataset corrupte by white Gaussian noise and noise standard deviation 25.
Overview
DCT2net software, based on the well-known DCT (Discrete Cosine Transform) image denoising algorithm, is dedicated to noise removal from images. The traditional DCT denoiser can be seen as a shallow CNN and thereby its original linear transformcan be tuned through gradient descent in a supervised manner, improving considerably its performance. Consequently, DCT2net is a shallow and interpretable convolution network, whose parameters optimization allows to improve very significantly the performances of the traditional DCT denoiser. To deal with the remaining artifacts induced by DCT2net, an original hybrid solution between DCT and DCT2net is proposed, combining the best of what these two methods can offer. Experiments on artificially noisy images show that the two layer DCT2netmethod provides results comparable to the BM3D method and is as fast as the DnCNN algorithm composed of more than a dozen of layers.
Software distribution
Work in progress
Reference
S. Herbreau, C. Kervrann. DCT2net: an interpretable shallow CNN for image denoising, arXiv-2107.14803, HAL-INRIA-03511641, 2021
