Online Turkish Handwriting Recognition Using Synthetic Data

Abstract

We present a recognition system for online Turkish handwriting trained with synthetically generated data and transfer learning. Training deep networks requires large amounts of data. However, a sufficiently large collection of Turkish handwriting samples is not available. Hence we synthesize data to do pretraining before adapting the system to target dataset by fine tuning. We generate words from isolated character collection of a large English handwriting dataset. Then, we train the system first with synthetic data and fine tune it with Turkish handwriting samples from a smaller dataset. Fine tuning increases the character recognition rate of the final system which is evaluated on 2,041 samples of isolated Turkish words from the initial value of 61% to 88%. Performance of the system on synthetic data is quite similar to that on the Turkish test data which shows that the synthetic data resembles the real data quite closely. According to these results, synthetic data generation can be a solution to the data scarcity problem of online Turkish handwriting.

Recognition of Online Turkish Handwriting using Transfer Learning

“…Taking an approach similar to the method proposed in this work, a CNN-BLSTM network, which is pre-trained with a synthetic dataset and fine-tuned with the ET train samples, achieved 12% CER and 44% WER on ET test set in [40]. Our results are approximately equal to those last results while we use a less complex method.…”

“…When the lexicon size is increased to 12,500, recognition accuracy is measured as 67.9%, using a bi-gram language model based on word stems and suffixes. In a recent study, a CNN-BLSTM network which is pre-trained with a synthetic dataset and fine-tuned with the Turkish dataset used in [24] achieved 88% character recognition accuracy in an open dictionary recognition task on that Turkish dataset [40].…”

Розпізнавання Рукописних Українських Літер Та Цифр З Використанням Синтетичного Набору Даних Та Згорткових Нейронних Мереж

“…Спочатку більшість досліджень була зосереджена на розпізнаванні літер латинського алфавіту, але в останні роки увагу привертають й інші алфавітиарабський, російський, казахський, китайський, індійський тощо [6][7][8][9][10][11]. Для досліджень технологій розпізнавання рукописних літер латинського алфавіту фактичним стандартом є набір даних EMNIST [12].…”