Random pixel permutation destroys local correlations in images, causing standard CNN classification accuracy to drop depending on class similarities while dilated convolutions recover some performance.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2019 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Convolutional Neural Networks on Randomized Data
Random pixel permutation destroys local correlations in images, causing standard CNN classification accuracy to drop depending on class similarities while dilated convolutions recover some performance.