Offline Handwritten Amharic Word Recognition using Deep Learning

Abstract

Amharic (Amarñña: አማርኛ) is the official language of the Federal Government of Ethiopia, with more than 27 million speakers. But it is a low-resourced language, and only a few attempts have been made so far for handwritten Amharic text recognition. This is challenging due to the very high similarity between many of the alphabets, and handwriting calligraphy is personal. This paper presents offline handwritten Amharic word recognition using deep learning architecture, which comprises convolutional neural networks (CNNs) for feature extraction from input word images, bidirectional recurrent neural networks (BRNNs) for sequence encoding, moreover connectionist temporal classification (CTC) as a loss function. To the authors knowledge, there was no any publicly available handwritten Amharic text dataset. Therefore, we commenced from preparing the dataset from scratch. We chose A* path-planning and scale space algorithms respectively for line and word level segmentation from the raw handwritten text image. We have collected 12 064 word-images for our dataset. Data augmentation was employed by applying random transformations to the word-images of the training set to enhance performance of the proposed models. In the main process, we have developed a custom model with CNN-BRNN-CTC framework, and Bayesian algorithm was used to set values for hyper-parameters. We have then compared four different state-of-the-art CNN models: EfficientNet, DenseNet, ResNet and VGG for robust feature extraction. We did this experiment by modifying their architectures to fit our problem domain. We have measured the word error rate (WER) and character error rate (CER) using our test set, which contains 1200 word images (they are randomly selected 10% of the total dataset). Hence, the outperforming model with DenseNet201 feature extractor network has achieved WER of 9.00 % and CER of 2.51 % on the non-augmented 64 × 256 word-image dataset and become advanced in WER of 6.83 % and CER of 1.82 % on augmented one. Whereas, the custom model has also achieved competitive performance on the offline handwritten Amharic word recognition with the state-of-the-art models