Single-Character-Based Embedding Feature Aggregation Using Cross-Attention for Scene Text Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Scene text super-resolution (STSR) aims to improve text quality for better readability and downstream tasks.
• Challenges in STSR include character ambiguity and interference from complex backgrounds, especially with tightly connected characters.

Purpose of the Study:

• To propose a single-character-based embedding feature aggregation method using cross-attention for scene text super-resolution (SCE-STISR).
• To address the challenges of character ambiguity and background interference in complex scene text images.

Main Methods:

• Employs a dynamic feature extraction mechanism with adaptive multi-scale feature weights.
• Introduces a dual-level cross-attention mechanism for aggregating single-character features with textual priors and aligning visual-semantic information.
• Applies adaptive normalized color correction to reduce background-induced color distortion.

Main Results:

• Achieved improved text recognition accuracies of 0.9-1.4% over the baseline TATT on the TextZoom benchmark.
• Obtained an optimal SSIM value of 0.7951 and a PSNR of 21.84 on TextZoom.
• Demonstrated accuracy improvements of 0.2-2.2% over existing baselines on five text recognition datasets.

Conclusions:

• The proposed SCE-STISR method effectively enhances scene text super-resolution by addressing character ambiguity and background interference.
• The approach validates the effectiveness of single-character embedding aggregation and cross-attention for improving text recognition accuracy in challenging scenarios.