CCDPlus: Towards Accurate Character to Character Distillation for Text Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Scene text recognition (STR) methods often rely on large-scale synthetic data (LSD), but a synth-to-real domain gap limits performance.
• Unlabeled real data (URD) holds valuable information, yet effectively utilizing it for STR remains challenging.
• Previous self-supervised methods on URD face issues like coarse representations, inflexible augmentation, and real-to-synth domain drift.

Purpose of the Study:

• To propose CCDPlus, a novel character-to-character distillation method for robust scene text recognition.
• To address the limitations of existing methods by integrating URD and LSD within a unified framework.
• To improve the efficiency and robustness of STR models in real-world scenarios.

Main Methods:

• CCDPlus employs a joint supervised and self-supervised learning framework for scene text recognition.
• It delineates fine-grained character structures on URD as representation units by transferring knowledge from LSD online.
• The method enables flexible character-to-character distillation with versatile data augmentation, extracting generalizable character-level features.

Main Results:

• CCDPlus achieves state-of-the-art (SOTA) performance, outperforming supervised, semi-supervised, and self-supervised methods.
• It demonstrates an average improvement of 1.8% over supervised, 0.6% over semi-supervised, and 1.1% over self-supervised methods on standard datasets.
• Significant improvement of 6.1% is observed on the challenging Union14M-L dataset.

Conclusions:

• CCDPlus effectively bridges the synth-to-real domain gap in scene text recognition.
• The character-to-character distillation approach enhances feature representation and recognition accuracy.
• The unified framework successfully combines self-supervised learning on URD with supervised learning on LSD for superior performance.