Language-Aware Non-Autoregressive Khmer Textline Recognition Using Khmer Subword Model

要約

Unlike the Latin script, Khmer does not use spaces between words, leading to text recognition typically being done at the textline level. This can involve a vast number of characters and results in high latency for a language-aware autoregressive (AR) decoder that generates one character at a time. On the other hand, a non-autoregressive (NAR) decoder generates all characters in parallel, but it is not language-aware. In this paper, we introduce an efficient Khmer textline recognition method based on a NAR decoder, ensuring low decoding latency while maintaining linguistic awareness. This is achieved by utilizing a Khmer-specific subword modeling called Khmer character clusters (KCC) that capture the syntactic, morphological, and orthographic aspects of the Khmer script. Therefore, instead of conventional character-level recognition, the proposed method recognizes all character clusters or subwords in parallel. The experimental results demonstrate that the proposed method outperforms the character-level baseline NAR model in terms of recognition accuracy while maintaining the same low latency. When compared with the character-level baseline AR model, the proposed method achieves comparable or improved recognition accuracy while also achieving significantly lower latency. When compared with the recent state-of-the-art (SOTA) NAR and AR Khmer text recognition methods, our proposed method achieves superior recognition performance.

関連ファイル

• …