Noisy-Aware Unsupervised Domain Adaptation for Scene Text Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation (UDA) is crucial for Scene Text Recognition (STR), enabling knowledge transfer from synthetic to real-world data.
• Current UDA methods for STR struggle with noisy pseudo-labels due to domain gaps (inter-domain noise) and environmental variations (intra-domain noise).

Purpose of the Study:

• To develop a novel noisy-aware unsupervised domain adaptation framework for Scene Text Recognition (STR).
• To enhance model robustness against both inter- and intra-domain noise for improved pseudo-label quality.
• To adapt the framework for the challenging Source-Free Unsupervised Domain Adaptation (SFUDA) setting.

Main Methods:

• Proposing a reweighting strategy for target pseudo-labels using refined probability distributions to mitigate domain gap impact.
• Introducing a decoupled triple-P-N consistency matching module with data augmentation for enhanced robustness in diverse environments.
• Implementing low-confidence character negative learning, decoupled from high-confidence positive learning, to improve sample utilization.

Main Results:

• The proposed framework significantly improves pseudo-label precision by addressing noise.
• Demonstrated superior performance and faster convergence in the Source-Free UDA (SFUDA) setting compared to existing UDA-based STR methods.
• Established new state-of-the-art results across multiple benchmark datasets for STR.

Conclusions:

• The noisy-aware UDA framework effectively enhances STR model robustness and performance.
• The method offers a significant advancement for UDA in STR, particularly in the challenging SFUDA scenario.
• The framework achieves state-of-the-art performance, highlighting its potential for real-world scene text recognition applications.