MIRROR: Multi-scale iterative refinement for robust chinese text recognition

Abstract

Text recognition has become a key area of research due to its wide applications in various fields. As an important branch of computer vision, Chinese text recognition has gained increasing research and practical value. However, the existing Chinese text recognition methods are still limited. This paper proposes an innovative Chinese text recognition method, Multi-Scale Iterative Refinement for Robust Chinese Text Recognition (MIRROR). The model significantly improves the recognition accuracy of Chinese text through advanced algorithms and structural design. The MIRROR model consists of two core components: a feature extractor and a Next-Character Decoder. Specifically, this paper proposes a Spatial Local Self-Attention Module to enhance the model’s ability to model long-distance dependencies in complex character sequences, addressing the problem of complex distributions in medium-to-long distance Chinese character sequences. The Character Refinement Module effectively captures multi-scale information, handles stroke feature differences, and resolves inter-class similarity issues. By combining multi-scale feature extraction with iterative optimization for feature refinement, the model identifies common features across different styles of the same character, solves the intra-class variation problem, and improves model robustness. In addition, this paper introduces a Three-Dimensional Weight Attention Module to refine the granularity of character features. Experiments show that MIRROR significantly outperforms baseline models on Chinese benchmark datasets. On scene datasets, performance improves by 3.08% (from 76.90% to 79.98%), on web datasets by 1.46% (from 70.43% to 71.89%), on document datasets by 0.38% (from 98.72% to 99.10%), and on handwriting datasets by 9.29% (from 50.26% to 59.55%).