Scale and orientation invariant text segmentation for born-digital compound images

Area of Science:

• Computer Vision
• Image Processing
• Document Analysis

Background:

• Born-digital compound images present challenges for text segmentation due to mixed textual and pictorial content.
• Accurate text segmentation is crucial for numerous downstream applications, including information retrieval and content analysis.

Purpose of the Study:

• To propose a robust coarse-to-fine framework for segmenting text from born-digital compound images.
• To handle texts of arbitrary scales and orientations effectively.
• To improve the precision and integrity of text segmentation.

Main Methods:

• A coarse stage utilizes a local image activity measure based on character variation distribution to differentiate textual and pictorial regions.
• A fine stage refines the segmentation using textual connected components (TCCs) and a scale/orientation invariant grouping algorithm.
• String-level features (shapeness, color similarity, mean activity) are employed to eliminate false positives.

Main Results:

• The framework accurately segments textual regions from complex born-digital compound images.
• Preserves the integrity of texts with diverse scales and orientations.
• Successfully avoids over-segmentation or merging of textual regions.

Conclusions:

• The proposed coarse-to-fine framework offers a precise and effective solution for text segmentation in born-digital compound images.
• The method demonstrates robustness in handling variations in text scale and orientation.
• This approach advances the capabilities of image analysis for documents containing mixed content.