Deconvolution and Checkerboard Artifacts
Abstract
Distill articles are interactive publications and do not include traditional abstracts. This summary was written for the ML Anthology. Identifies the source of checkerboard artifacts commonly found in neural network-generated images, tracing them to uneven overlap in deconvolutional layers. Proposes resize-convolution as a simple alternative that eliminates these artifacts.
Cite
Text
Odena et al. "Deconvolution and Checkerboard Artifacts." Distill, 2016. doi:10.23915/distill.00003Markdown
[Odena et al. "Deconvolution and Checkerboard Artifacts." Distill, 2016.](https://mlanthology.org/distill/2016/odena2016distill-deconvolution/) doi:10.23915/distill.00003BibTeX
@article{odena2016distill-deconvolution,
title = {{Deconvolution and Checkerboard Artifacts}},
author = {Odena, Augustus and Dumoulin, Vincent and Olah, Chris},
journal = {Distill},
year = {2016},
doi = {10.23915/distill.00003},
url = {https://mlanthology.org/distill/2016/odena2016distill-deconvolution/}
}