Pull Pole Points to Text Contour by Magnetism: A Real-Time Scene Text Detector

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Scene text detection is vital for scene understanding and has seen significant research interest.
• Segmentation-based methods offer adaptable pixel-level predictions but face challenges with contour reconstruction.
• Existing methods like shrink masks and Vatti clipping have limitations in capturing local details and efficiency.

Purpose of the Study:

• To propose an efficient and effective scene text detection method, the Magnetic Text Detector (MTD).
• To address limitations of existing methods regarding local contour information and post-processing complexity.
• To improve both the accuracy and efficiency of scene text detection.

Main Methods:

• Introduced a novel text representation: flexible mask (FM), which adapts shrinkage based on local contours.
• Developed a magnetic pull module (MPM) that uses generated magnetic fields to guide FM pole points to text contours.
• Proposed Magnetic Text Detector (MTD) integrating FM and MPM for robust contour reconstruction.

Main Results:

• The flexible mask (FM) preserves local contour information and text instance distinctiveness.
• The magnetic pull module (MPM) enables accurate contour reconstruction even with initial prediction deviations.
• MTD approximately saves 50% of post-processing time compared to existing methods.
• MTD achieves state-of-the-art (SOTA) performance on multiple scene text detection datasets.

Conclusions:

• The proposed Magnetic Text Detector (MTD) effectively enhances scene text detection accuracy and efficiency.
• The flexible mask (FM) and magnetic pull module (MPM) are key components contributing to MTD's success.
• MTD offers a promising solution for scene text detection challenges, outperforming existing SOTA methods.