Repository of Image Databases: Distilled Images

Training a CNN with a smaller set of distilled images provides classification statistics with a modest differences versus the statistics obtained after training with the entire original training set.

Visual Validation

Downloads

CIFAR-10

• 100 images (with 10 per class) distilled from CIFAR-10 applying the M-PCA-SVD method (a non-gradient and non-neural network method) *
  • Code for this method can be downloaded at this link. *
• 100 images (with 10 per class) distilled from CIFAR-10 applying – ResNet CNN with GALER method (Math 595 – Summer 2026) *
  • Trained the same CNN and made it obtain 76.42% of accuracy after 1000 epochs.

Digit-MNIST

• 100 images distilled from the digit-MNIST database (with 10 per class). *
• 100 images (with 10 per class) distilled from digit-MNIST applying – LeNet CNN with GALER method (Math 595 – Summer 2026) *
  • Trained the same CNN and obtain 99.28% of accuracy after 1000 epochs.

OCID

• 20 images distilled from the training OCID database, with 224 x 224 images, after 200 training epochs. *

Results are shown in Table 5 of paper at https://doi.org/10.3390/math13233785

*The downloadable ZIP file (compressed folder) containing multiple images and data files; users will need to download and extract the files to access the contents.

Normal (OCID)

OSCC (OCID)