Real-Time Scene Text Detection With Differentiable Binarization and Adaptive Scale Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Segmentation-based methods excel at detecting arbitrary-shaped text but suffer from complex post-processing and scale robustness issues.
• Existing approaches often involve time-consuming, isolated post-processing steps and rudimentary multi-scale feature fusion.
• These limitations hinder both the optimization and efficiency of scene text detection models.

Purpose of the Study:

• To develop a more accurate and efficient scene text detection method.
• To address the limitations of complex post-processing and scale robustness in current segmentation-based approaches.
• To integrate key post-processing steps into the network for end-to-end optimization.

Main Methods:

• Proposed a Differentiable Binarization (DB) module, integrating the binarization process into the segmentation network.
• Introduced an Adaptive Scale Fusion (ASF) module for improved scale robustness by adaptively fusing multi-scale features.
• Combined the DB and ASF modules with a segmentation network for a streamlined scene text detection pipeline.

Main Results:

• The integrated DB module optimizes the segmentation network for enhanced text detection accuracy.
• The ASF module effectively improves the model's robustness across different text scales.
• The proposed scene text detector achieved state-of-the-art performance in both accuracy and speed on five benchmarks.

Conclusions:

• The proposed Differentiable Binarization and Adaptive Scale Fusion modules significantly improve scene text detection.
• Integrating binarization into the network simplifies the pipeline and boosts accuracy.
• The method offers a superior balance of accuracy and speed for real-world scene text detection applications.