Deep learning based offline handwritten character recognizer systems with a multilingual handwritten character dataset

Despite decades of research, offline handwriting recognition is still an unresolved research problem. Advancements in deep learning led to a boost in image processing domain in general including the recognition of offline handwritings. In order to max out the capabilities of deep learning-based methods, a large input set is essential. However, these is a lack of publicly available handwriting datasets, especially in certain languages. Absence of handwritten character datasets in Turkish and Hungarian was a prompt to create a handwritten character dataset in those languages. In this work, a public domain multilingual handwritten character dataset is generated. In addition to the proposed T-H-E Dataset, two different offline multilingual handwriting recognition systems were developed. The first one is a segmentation-based recognizer, using a novel deep learning architecture put forward in this study. In attempt to create a larger input for the network, synthetic characters based on the characters in TH-E Datasets are generated. Deep Convolutional Generative Adversarial Networks (DCGANs) are adopted to create synthetic data for augmenting the existing dataset. Additionally, a segmentation-free handwriting recognizer is proposed as the second recognizer. The latest version of YOLO network for object detection, namely YOLOv5 is applied to the system. An object detection algorithm takes a large image containing one or multiple objects as an input and predicts the location and class label of the objects. Based on this assumption, the handwritten characters are systematically placed onto a 416×416-pixel image thus creating a suitable input to YOLOv5 network. The results indicate that a segmentation-based recognition is ideal for the recognition of non-cursive single language handwritten texts. However, object recognition-based system outperforms the segmentation-based method in both single language and multilingual handwritten text recognition for cursive and non-cursive characters.