Enhancing Scene Text Detection with Parameter-Free Attention Mechanism and an Improved FPN

In scene text detection, models like DBNet excel in handling texts of arbitrary shapes, but there is still potential for improvement in detecting small texts. This study introduces two key enhancements to the DBNet framework: first, a Simple and Parameter-Free Attention Module that sharpens the network's focus on important features while diminishing the impact of complex backgrounds, thus improving detection accuracy for small text areas. Second, a Cross-Level Enhanced Feature Fusion Module is proposed to integrate features across different scales, effectively utilizing local and global contextual information. This multi-scale fusion significantly boosts the model's capability to extract features of small texts, ensuring precise localization. Comprehensive experimental results demonstrate that these improvements notably enhance the model's performance across various challenging datasets, especially in detecting small text objects, showcasing superior accuracy and robustness.