Reproduce “Efficient scene text image super-resolution with semantic guidance” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
LeoWu TomyEnrique, Xiangcheng Du, Kangliang Liu, Han Yuan, Zhao Zhou, Cheng Jin
SGENet, an efficient framework for Scene Text Image Super-Resolution (STISR), leverages a lightweight text recognizer for semantic guidance and a visual-semantic alignment module. It achieves competitive text recognition accuracy, reducing computational complexity by 85% and parameters by 78% compared to state-of-the-art methods, making it suitable for resource-limited deployments.
Key insights
- Effective semantic guidance for STISR can be achieved using lightweight pre-trained text recognizers, avoiding the prohibitive computational costs of larger models while maintaining strong performance.
- A bidirectional cross-attention mechanism is efficient for aligning semantic text features with visual image features, providing rich contextual information for text reconstruction.
- Careful design choices, such as optimizing the number of Sequential Recurrent Blocks (SRBs), are critical for achieving an optimal balance between performance and computational efficiency in STISR architectures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.