GLaLT: Global-Local Attention-Augmented Light Transformer for Scene Text Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Scene Text Recognition (STR) benefits from both Connectionist Temporal Classification (CTC) and attention mechanisms.
• CTC offers computational efficiency but lacks effectiveness, while attention mechanisms are effective but computationally intensive.
• There is a need for methods that balance speed and accuracy in STR.

Purpose of the Study:

• To propose a novel architecture, the global-local attention-augmented light Transformer (GLaLT), for scene text recognition.
• To integrate the strengths of CTC and attention mechanisms for improved STR performance.
• To achieve high accuracy and computational efficiency simultaneously in STR.

Main Methods:

• GLaLT employs a Transformer-based encoder-decoder structure.
• The encoder fuses self-attention for global dependencies and convolution for local context.
• The decoder utilizes parallel attention and CTC modules, with attention guiding CTC during training.

Main Results:

• GLaLT achieves state-of-the-art performance on standard benchmarks for both regular and irregular scene text recognition.
• The model demonstrates a favorable trade-off between speed, accuracy, and computational efficiency.
• GLaLT effectively balances the effectiveness of attention with the efficiency of CTC.

Conclusions:

• GLaLT represents a significant advancement in scene text recognition by synergistically combining CTC and attention.
• The proposed architecture offers a practical solution for real-world STR applications demanding both speed and accuracy.
• GLaLT pushes the boundaries for maximizing speed, accuracy, and computational efficiency in STR.