Lightweight Scene Text Recognition Based on Transformer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Scene text recognition (STR) is crucial for computer vision applications.
• Current attention-based models face challenges with attention drift in complex/low-quality images.
• Transformer models increase computational costs due to large parameter counts.

Purpose of the Study:

• To address attention drift and high computational costs in STR.
• To improve recognition accuracy and efficiency in challenging visual scenes.
• To develop a robust STR model balancing performance and resource usage.

Main Methods:

• Utilized Vision Transformer (ViT) architecture.
• Introduced an additional position-enhancement branch to mitigate attention drift.
• Dynamically fused positional and visual information for improved feature representation.

Main Results:

• Achieved a 3% higher average recognition accuracy on the test set compared to the baseline model.
• Maintained a small parameter count and fast inference speed.
• Demonstrated a favorable balance between recognition accuracy, processing speed, and computational load.

Conclusions:

• The proposed ViT-based model effectively alleviates attention drift in STR.
• The dynamic fusion of information enhances recognition performance.
• The model offers an efficient and accurate solution for real-world scene text recognition challenges.