Attention-Based Scene Text Detection on Dual Feature Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Segmentation-based scene text detection excels with arbitrary shapes and extreme aspect ratios.
• Existing methods struggle with insufficient semantic and spatial information, leading to feature loss.
• Limited feature representation hinders accurate text classification and positioning.

Purpose of the Study:

• To propose an Attention-based Dual Feature Fusion Model (ADFM) for enhanced scene text detection.
• To address the limitations of insufficient semantic and spatial information in current algorithms.
• To improve the classification and positioning capabilities of scene text detection networks.

Main Methods:

• Introduced a Bi-directional Feature Fusion Pyramid Module (BFM) for enhanced multi-scale semantic information.
• Incorporated a position-sensitive Spatial Attention Module (SAM) to focus on relevant text features.
• Employed a two-stage feature fusion process with top-down and bottom-up pathways.

Main Results:

• The ADFM effectively enhances the representation of multi-scale text semantic information.
• The Spatial Attention Module improves the network's sensitivity to text regions.
• Ablation experiments validated the effectiveness of individual ADFM modules.

Conclusions:

• The proposed ADFM significantly improves scene text detection performance.
• The model demonstrates superior capabilities in handling complex text scenarios.
• ADFM offers a promising approach for advanced scene text detection tasks.