When Image Denoising Meets High-Level Vision Tasks: A Deep Learning Approach

Authors: Ding Liu, Bihan Wen, Xianming Liu, Zhangyang Wang, Thomas Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate their denoising performance over the widely used Kodak dataset2, which consists of 24 color images. Table ?? shows the peak signalto-noise ratio (PSNR) results for CBM3D [Dabov et al., 2007a], TNRD [Chen and Pock, 2017], MCWNNM [Xu et al., 2017], Dn CNN [Zhang et al., 2017a], and our proposed method. Table ?? and Table ?? list the performance of high-level vision tasks, i.e., top-1 and top5 accuracy for classification and mean intersection-over-union (Io U) without conditional random field (CRF) postprocessing for semantic segmentation.
Researcher Affiliation Collaboration 1 University of Illinois at Urbana-Champaign, USA 2 Facebook Inc. 3 Texas A&M University, USA
Pseudocode No The paper provides network architectures in diagrams but no pseudocode or algorithm blocks.
Open Source Code Yes The code is available online1. 1https://github.com/Ding-Liu/Deep Denoising
Open Datasets Yes We use the training set as in [Chen et al., 2014]. We use the training set as in [Chen et al., 2014]. We evaluate their denoising performance over the widely used Kodak dataset2, which consists of 24 color images. 2http://r0k.us/graphics/kodak/. we train our model on ILSVRC2012 training set. we train our model on the augmented training set of Pascal VOC 2012 as in [Chen et al., 2014].
Dataset Splits Yes evaluate the classification accuracy on ILSVRC2012 validation set. test on its validation set.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) are mentioned.
Software Dependencies No The paper mentions VGG-based networks and SGD, but no specific software versions for libraries or frameworks are provided.
Experiment Setup Yes We use SGD with a batch size of 32, and the input patches are 48 48 pixels. The initial learning rate is set as 10 4 and is divided by 10 after every 500,000 iterations. The training is terminated after 1,500,000 iterations. λ is empirically set as 0.25. λ is empirically set as 0.5.