HGR-Net: Hierarchical Graph Reasoning Network for Arbitrary Shape Scene Text Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Scene text detection is crucial for text reading but remains challenging due to text variability.
• Existing methods often use bottom-up component extraction or top-down contour regression.
• A hybrid approach is needed to handle diverse text shapes and orientations.

Purpose of the Study:

• To propose a novel Hierarchical Graph Reasoning Network (HGR-Net) for arbitrarily-shaped scene text detection.
• To develop a method that integrates both instance-level and component-level text representations.
• To improve the accuracy and robustness of scene text detection.

Main Methods:

• Introduced HGR-Net with a Text Feature Extraction Network (TFEN) and a Text Relation Learner Network (TRLN).
• TFEN extracts multi-grained text candidates (instance-level contours, component-level quadrangles).
• TRLN uses inter-text and intra-text graphs for contextual and geometric reasoning, bridging hierarchical levels.

Main Results:

• HGR-Net demonstrated state-of-the-art performance on arbitrary orientation and shape scene text detection.
• Achieved superior results on SCUT-CTW1500, Total-Text, and ICDAR15 benchmarks.
• The hierarchical relational reasoning effectively leverages complementary graph embeddings.

Conclusions:

• HGR-Net offers a powerful hybrid approach for complex scene text detection.
• The proposed hierarchical graph reasoning effectively captures intricate text structures.
• This method advances the state-of-the-art in arbitrary shape scene text detection.