Stroke-Based Scene Text Erasing Using Synthetic Data for Training.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Scene text erasing is crucial for image editing and privacy.
• Existing methods struggle due to a lack of large-scale real-world datasets for text removal.
• Text detection and image inpainting are key subtasks requiring substantial data.

Purpose of the Study:

• To develop an effective scene text erasing method despite the scarcity of real-world data.
• To propose a novel network architecture for accurate text stroke extraction and background inpainting.
• To enable automatic scene text erasing by integrating with existing text detectors.

Main Methods:

• Utilized an enhanced synthetic text engine to generate a large-scale training dataset.
• Developed a network with a stroke mask prediction module and a background inpainting module.
• Extracted text strokes as small holes to preserve background details for improved inpainting.

Main Results:

• The proposed method significantly outperforms state-of-the-art methods on SCUT-Syn, ICDAR2013, and SCUT-EnsText datasets.
• Achieved superior performance even when compared to methods trained on real-world data.
• Demonstrated the effectiveness of using enhanced synthetic data for training.

Conclusions:

• The proposed scene text erasing method is highly effective, even without real-world training data.
• The novel network architecture successfully handles text removal and background reconstruction.
• This approach offers a viable solution for automatic and partial scene text erasing.