MR-FPN: Multi-Level Residual Feature Pyramid Text Detection Network Based on Self-Attention Environment

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Text detection in complex backgrounds remains a significant challenge in the age of intelligence.
• Existing algorithms often lack robustness, struggle with locating text regions, and misdetect adjacent text instances.

Purpose of the Study:

• To propose a robust text detection method capable of accurately separating adjacent text instances in complex backgrounds.
• To enhance feature representation and fusion for improved text detection accuracy and robustness.

Main Methods:

• Developed a Multi-Level Residual Feature Pyramid Network (MR-FPN) using ResNet50 as the backbone.
• Integrated a Self-Attention Module (SAM) to capture pixel-level relations and contextual information.
• Incorporated a Multi-Scale Enhancement Module (MEM) to improve text information expression and semantic extraction.

Main Results:

• The MR-FPN model demonstrated improved feature fusion and provided refined features for text detection.
• Evaluated on multiple datasets (CTW1500, Total-Text, ICDAR2015, MSRA-TD500), the model achieved significant improvements.
• Achieved an F-measure of 83.31% on the Total-Text dataset, surpassing the baseline by 5%.

Conclusions:

• The proposed MR-FPN effectively enhances the robustness of text detection in challenging environments.
• The integration of self-attention and multi-scale enhancement modules significantly boosts performance.
• The model shows strong potential for real-world applications requiring accurate text detection.