Scale-Wise Convolution for Image Restoration

Authors: Yuchen Fan, Jiahui Yu, Ding Liu, Thomas S. Huang10770-10777

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments, we compare the restoration accuracy and parameter efficiency among our model and many different variants of multi-scale neural networks. The proposed network with scale-wise convolution achieves superior performance in multiple image restoration tasks including image super-resolution, image denoising and image compression artifacts removal.
Researcher Affiliation Collaboration Yuchen Fan,1 Jiahui Yu,1 Ding Liu,2 Thomas S. Huang1 1University of Illinois at Urbana-Champaign, 2Bytedance Inc.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks. The model architecture is described textually and illustrated with a diagram (Figure 2).
Open Source Code Yes Code and models are available at: https://github.com/ychfan/scn sr.
Open Datasets Yes For image super-resolution, models are trained on DIV2K (Timofte et al. 2017) dataset... For image denoising, we use Berkeley Segmentation Dataset (BSD) (Martin et al. 2001)... For compression artifacts removal, we use 91 images in (Yang et al. 2010) and 200 training images in (Martin et al. 2001).
Dataset Splits Yes The default splits of DIV2K dataset consist 800 training images and 100 validation images.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or other computer specifications used for running the experiments.
Software Dependencies No The paper mentions 'ADAM optimizer' but does not provide specific software dependencies or library names with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA versions) needed to replicate the experiment.
Experiment Setup Yes Data augmentation including flipping and rotation are performed online during training, and Gaussian noise for image denoising tasks is also online sampled. 1000 patches randomly sample per image and per epoch, and 40 epochs in total. All models are trained with L1 distance through ADAM optimizer, and learning rate starts from 0.001 and halves every 3 epochs after the 25th. We use deep models with 8 residual blocks, 32 residual units and 4x width multiplier for most experiments, and 64 blocks and 64 units for super-resolution benchmarks.