Training deep learning based denoisers without ground truth data

Authors: Shakarim Soltanayev, Se Young Chun

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, denoising simulation results are presented with the MNIST dataset using a simple stacked denoising autoencoder (SDA), and a large-scale natural image dataset using a deep convolutional neural network (CNN) image denoiser (Dn CNN).
Researcher Affiliation Academia Shakarim Soltanayev Se Young Chun Department of Electrical Engineering Ulsan National Institute of Science and Technology (UNIST), Republic of Korea {shakarim,sychun}@unist.ac.kr
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/Shakarim94/Net-SURE.
Open Datasets Yes We performed denoising simulations with the MNIST dataset. The network was trained with 400 images with matrix sizes of 180 180 pixels. Two test sets were used to evaluate performance: one set consisted of 12 widely used images (Set12) [17], and the other was a BSD68 dataset.
Dataset Splits Yes SDA was trained to output a denoised image using a set of 55,000 training and 5,000 validation images.
Hardware Specification Yes With the use of an NVidia Titan X GPU, the training process took approximately 7 hours for Dn CNNMSE-GT and approximately 11 hours for Dn CNN-SURE.
Software Dependencies No The paper mentions 'Tensor Flow [34]' as a deep learning development framework but does not specify a version number.
Experiment Setup Yes For all cases, SDA was trained with the Adam optimization algorithm [33] with the learning rate of 0.001 for 100 epochs. The batch size was set to 200 (bigger batch sizes did not improve the performance). The ϵ value in (6) was set to 0.0001.