Searching a High Performance Feature Extractor for Text Recognition Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Feature extractors are crucial for text recognition (TR) but manual architecture customization is complex and time-consuming.
• Neural Architecture Search (NAS) has shown success in automating model design, but existing methods are insufficient for complex TR feature extractor search spaces.

Purpose of the Study:

• To develop an automated method for discovering optimal feature extractor architectures for text recognition.
• To design a domain-specific search space tailored for feature extractors in TR.
• To propose an effective NAS algorithm capable of navigating a large and complex search space.

Main Methods:

• Designed a domain-specific search space encompassing 3D spatial and transformed-based sequential models.
• Developed a two-stage NAS algorithm: progressive block training with an auxiliary head, followed by latency-constrained sub-network search using natural gradient descent.
• Conducted ablation studies to validate the search space, algorithm, and resulting architectures.

Main Results:

• The proposed NAS method effectively searches a large and complex feature extractor space.
• Searched architectures achieve state-of-the-art performance on both handwritten and scene text recognition tasks.
• The method demonstrates improved recognition accuracy while significantly reducing computational latency.

Conclusions:

• Automated feature extractor search using NAS is a viable and effective approach for text recognition.
• The proposed two-stage search algorithm and domain-specific search space enable efficient discovery of high-performance TR models.
• This work offers a practical solution for optimizing TR systems with reduced latency and enhanced accuracy.