for the training of the network, simply some parameters were adjusted in the part of data increase, like increasing zoom range, disabling flip and rotations.
other than that, a little digital image processing to rotate the images, data structures in python to extract and export the data, a little bit of programming
for use this dataset, you need to download the data and extract the data and run this codes
"cifar10_cnn.py" train the network
"data_pipe.py" have a util function to load the data for training
"fix_data.py" get the trained model and fix the dataset
possible improvements not tested
add some color changing in the data augmentation (HUE, gamma correction, etc)