Text-Line Detection in Camera-Captured Document Images Using the State Estimation of Connected Components.

Area of Science:

• Computer Vision
• Document Image Analysis

Background:

• Camera-based text processing is gaining traction.
• Existing methods primarily focus on scene text detection, neglecting camera-captured documents.
• Document understanding requires accurate text-line detection.

Purpose of the Study:

• To develop a robust text-line detection algorithm for camera-captured document images.
• To improve the foundational step for document understanding from images.
• To address limitations of existing methods in handling document images.

Main Methods:

• A connected component (CC)-based framework incorporating state estimation (scale and orientation) is proposed.
• Maximally Stable Extremal Regions (MSER) algorithm extracts CCs.
• Projection profiles are used for estimating CC scales and orientations, guiding a bottom-up clustering process.

Main Results:

• The algorithm effectively handles arbitrarily oriented text lines and various scales.
• A text-line/non-text-line classifier filters out background noise.
• The proposed method demonstrates superior performance compared to conventional approaches on standard datasets.

Conclusions:

• The developed text-line detection algorithm is effective for camera-captured documents.
• The integration of state estimation improves robustness and accuracy.
• The method shows promise for challenging datasets and advancing document image analysis.