Explicitly-Decoupled Text Transfer With Minimized Background Reconstruction for Scene Text Editing.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Scene text editing is crucial for applications like data generation and privacy protection.
• Existing methods often fuse text and background, leading to suboptimal results.
• There is a need for methods that preserve background integrity during text replacement.

Purpose of the Study:

• To propose a novel Scene Text Editing network, STEEM.
• To decouple text style and content from the background.
• To minimize background reconstruction for improved realism.

Main Methods:

• STEEM employs a text-background separation module to isolate source text.
• A style-guided text transfer module replaces source text with target text.
• A context-focused background reconstruction module refines the edited image, minimizing background alterations.

Main Results:

• STEEM achieved a lower FID index (24.67 vs. 29.48), indicating improved image quality.
• Text recognition accuracy increased from 76.8% to 78.8% with STEEM.
• Experimental evaluations on two datasets confirmed the effectiveness of the proposed approach.

Conclusions:

• STEEM offers a superior approach to scene text editing by decoupling text and background.
• The method effectively preserves background details while accurately replacing text.
• STEEM represents a significant advancement in scene text editing technology.