Multilingual Scene Text Detection Using Gradient Morphology

Literature Survey

Current methodologies for scene text detection are based on detecting stable homogeneous components having near similar intensities and refining them on the basis of their stroke properties. Morphological operations are also utilized in some of these works to eliminate spurious components and retaining maximum text components.

The concept of detecting Maximally Stable Extremal Regions (MSER) was first introduced by Matas et al. (2004) where stable blobs of text regions having similar intensity values are detected. Epshtein et al. (2010) introduced the concept of Stroke Width Transform (SWT) where a text component is considered to be a well-defined stroke depicting some meaning. Yao et al. (2012) uses this concept to localize multi-oriented text regions from complex natural scenes. Text components are retained using stroke properties of candidate components. Since MSER method is highly sensitive to blur, Chen et al. (2011) designed a Canny edge based MSER technique and applied SWT to discard the non-text MSER components. This combination of MSER and SWT has become a popular way of identifying text candidates. Gomez & Karatzas (2013) used this combination to detect multi-lingual scene texts and performing perceptual clustering to localize the text regions. Yin et al. (2015) extracted MSER blobs and applied morphology-based grouping to localize multi-oriented and multi-lingual scene texts.