SLOAN: Scale-Adaptive Orientation Attention Network for Scene Text Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks have advanced scene text recognition.
• Existing methods struggle with arbitrary-orientation text and single-scale features.
• Effective context modeling for varied character scales is challenging.

Purpose of the Study:

• To propose a novel scale-adaptive orientation attention network for arbitrary-orientation scene text recognition.
• To address limitations in handling diverse text scales and orientations.
• To improve context modeling for scene text recognition.

Main Methods:

• A dynamic log-polar transformer converts arbitrary rotations/scales into log-polar space shifts.
• A sequence recognition network with a character-level receptive field attention module encodes contexts.
• The entire architecture is trained end-to-end.

Main Results:

• The proposed network generates rotation-aware and scale-aware visual representations.
• The character-level attention module effectively encodes contexts for various-scale characters.
• Experiments on public datasets demonstrate superior performance.

Conclusions:

• The novel scale-adaptive orientation attention network significantly enhances arbitrary-orientation scene text recognition.
• The dynamic log-polar transformer and attention module are key to improved performance.
• The method offers a robust solution for real-world scene text recognition challenges.