A Multi-Scale Natural Scene Text Detection Method Based on Attention Feature Extraction and Cascade Feature Fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Scene text detection is vital for applications but struggles with diverse text instances and complex backgrounds.
• Existing methods often lack robustness in capturing varying text scales, shapes, and environmental complexities.

Purpose of the Study:

• To propose a robust multi-scale scene text detection method for natural scenes.
• To improve feature extraction and fusion for enhanced text detection capabilities.

Main Methods:

• Developed a multi-scale method incorporating an attention feature fusion module (DSAF) for global and local attention.
• Utilized an improved cascaded feature fusion module (PFFM) to integrate feature maps and expand receptive fields.
• Introduced a lightweight subspace attention module (SAM) for feature map partitioning and spatial information interaction.

Main Results:

• The proposed method demonstrated superior performance on ICDAR2015, Total-Text, and MSRA-TD500 datasets.
• Achieved significant improvements in accuracy, recall, and F-score compared to existing methods.
• Validated the effectiveness and practicality of the attention-based multi-scale approach.

Conclusions:

• The proposed method effectively addresses the limitations of current scene text detection techniques.
• The integration of attention mechanisms and cascaded feature fusion enhances the detection of diverse text instances.
• The method offers a practical and effective solution for real-world scene text detection challenges.