Robust Image Denoising of No-Flash Images Guided by Consistent Flash Images

Authors: Geunwoo Oh, Jonghee Back, Jae-Pil Heo, Bochang Moon

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that our technique can produce more robust results than state-of-the-art methods, given various flash/no-flash pairs with inconsistent image structures. The source code is available at https://github.com/CGLab-GIST/RIDFn F.
Researcher Affiliation Academia 1Gwangju Institute of Science and Technology, South Korea 2Sungkyunkwan University, South Korea {gnuo8325, jongheeback}@gm.gist.ac.kr, jaepilheo@skku.edu, bmoon@gist.ac.kr
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes The source code is available at https://github.com/CGLab-GIST/RIDFn F.
Open Datasets Yes We have trained the learning-based methods (FFDNet, Uformer-B, Restormer, DJF, CUNet, DC, and ours) using the public dataset (Aksoy et al. 2018), which includes 2775 flash/no-flash image pairs categorized into six classes: people, shelves, plants, toys, rooms, and objects.
Dataset Splits Yes Specifically, we have randomly divided the data set into three sets: 2263 images for the training, 256 images for the validation, and the other 256 images for the test.
Hardware Specification Yes Then, we have trained our neural network using Adam optimizer (Kingma and Ba 2015) for 50 epochs, and it has taken 72 hours given two NVIDIA TITAN RTX graphics cards.
Software Dependencies No The paper mentions software components like 'Adam optimizer' and 'Rectified Linear Unit (ReLU)' activation function, but does not provide specific version numbers for any programming languages, libraries, or frameworks used (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes Precisely, the network consists of nine convolutional layers, and each of which uses 80 filters of size 5x5, except for the last one that uses KxK + |Ωc| filters to generate the kc and wci, respectively... The sizes of the neighboring pixels Ωc and convolutional kernel kc are set to 15x15 and 7x7, respectively... The initial learning rate has been set to 0.0005 and reduced to 0.0001 after 15 epochs. We have used 64x64 image patches and set the batch size to 64.