Efficient Chinese Text Recognition Model Utilizing Feature Detail Attention

Text recognition is a research topic in computer vision, with limited achievements in Chinese text recognition. The intricate font structures and extensive character categories inherent to Chinese characters pose significant challenges for recognition systems. To address these challenges, this paper proposes a new single visual Chinese recognition model called FUD. Initially, a Position Encoding Correction Module (PECM) is developed to rectify and augment the positional information within the features. Then, to capture the intricate Chinese structures, a local-global fusion attention computation approach is proposed. In addition, a CLA mechanism is introduced to estimate the inductive bias advantages. This model effectively concentrates on both fine-grained feature details and the global information in the input image. The proposed method outperforms the existing methods by an impressive accuracy of 74.4% on a benchmark dataset for Chinese scene text recognition, with only 44 million parameters.