Transfer Learning Using Rotated Image Data to Improve Deep Neural Network Performance

In this work we explore the idea that, in the presence of a small training set of images, it could be beneficial to use that set itself to obtain a transformed training set (by performing a random rotation on each sample), train a source network using the transformed data, then retrain the source network using the original data. Applying this transfer learning technique to three different types of character data, we achieve average relative improvements between 6% and 16% in the classification test error. Furthermore, we show that it is possible to achieve relative improvements between 8% and 42% in cases where the amount of original training samples is very limited (30 samples per class), by introducing not just one rotation but several random rotations per sample.