Research article on IEEE of the topic: "Toward a Low-Resource Non-Latin-Complete Baseline: An Exploration of Khmer Optical Character Recognition"
by RINA BUOY 1 , (Graduate Student Member, IEEE), MASAKAZU IWAMURA 1 , (Member, IEEE), SOVILA SRUN2 , AND KOICHI KISE1
Many existing text recognition methods rely on the structure of Latin characters and words. Such methods may not be able to deal with non-Latin scripts that have highly complex features, such as character stacking, diacritics, ligatures, non-uniform character widths, and writing without explicit word boundaries. In addition, from a natural language processing (NLP) perspective, most non-Latin languages are considered low-resource due to the scarcity of large-scale data. This paper presents a convolutional Transformer-based text recognition method for low-resource non-Latin scripts, which uses local two dimensional (2D) feature maps. The proposed method can handle images of arbitrarily long text lines, which may occur with non-Latin writing without explicit word boundaries, without resizing them to a fixed size by using an improved image chunking and merging strategy. It has a low time complexity in self-attention layers and allows efficient training. The Khmer script is used as the representative of non-Latin scripts because it shares many features with other non-Latin scripts, which makes the construction of an optical character recognition (OCR) method for Khmer as hard as that for other non-Latin scripts. Thus, by analogy with the AI-complete concept, a Khmer OCR method can be considered as one of the non-Latin-complete methods and can be used as a low-resource non-Latin baseline method. The proposed 2D method was trained on synthetic datasets and outperformed the baseline models on both synthetic and real datasets. Fine-tuning experiments using Khmer handwritten palm leaf manuscripts and other non-Latin scripts demonstrated the feasibility of transfer learning from the Khmer OCR method. To contribute to the low-resource language community, the training and evaluation datasets will be made publicly available.