Year: 2025
Type: article
Abstract: Text detection plays a vital role in applications like automated document analysis and scene understanding, yet achieving reliable accuracy in cluttered or low-contrast environments remains challenging. We propose FAF-Text, an English text detection framework that integrates adaptive feature filtering and multi-scale fusion to address these limitations. The filtering module employs gradient analysis to suppress noise and irrelevant patterns, while the fusion mechanism dynamically combines contextual and semantic features through attention-based learning. Evaluations on benchmark datasets demonstrate a 23% improvement in edge preservation and 18% enhancement in multi-scale recognition compared to existing methods. Ablation studies confirm the necessity of both modules, particularly under high-noise and low-resolution conditions, Furthermore, the framework's modular architecture ensures compatibility with multilingual OCR systems, offering a balance between computational efficiency and adaptability to complex text layouts.
Authors: Ruixia Huang, Yunpeng Ji
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4413005827