method: CNN-WRDF2016-11-29

Authors: Wei Li

Description: Description: This is a single CNN with recycled deep features (WRDF, for recognizing text strings. The CNN is trained only using synthetic images with random strings. No dictionary, lexicon or language model is involved. Its performance is expected to be incomparable with those models trained on normal text samples.
Images with single characters or digits were excluded from training. Therefore, single-character and single-digit test images are omitted.