Article: Scene text detection using robust masks and cascaded classifiers Journal: International Journal of Computational Science and Engineering (IJCSE) 2025 Vol.28 No.3 pp.255 - 266 Abstract: Detecting text in scenes poses a significant and challenging problem due to the complex character shapes and the diverse degradations present in natural images. It represents the initial and crucial phase that must be successfully completed before the text recognition stage. In this study, we suggest a hybrid approach to tackle this issue, leveraging the maximally stable extremal regions (MSER) algorithm, which gained significant attention in recent research. Despite its popularity, it remains very sensitive to the shape, size, scale and background noise of text characters. To tackle its limits and refine the final result, we focus on an extended MSER based method. Overall flowchart of the suggested system is divided into three steps: 1) robust masks generation to identify the text candidate regions; 2) feature extraction and selection based on VGG16 deep learning architecture; 3) employing classifiers in a cascaded structure to differentiate between text and non-text areas based on enhanced geometrical pattern characteristics. The effectiveness of the proposed method is demonstrated through an experimental study conducted on various benchmarks, such as ICDAR2013, ICDAR2015, MSRA-TD500, and RRC-MLT. Inderscience Publishers - linking academia, business and industry through research

Title: Scene text detection using robust masks and cascaded classifiers

Authors: Houssem Turki; Mohamed Elleuch; Monji Kherallah; Alima Damak

Addresses: National Engineering School of Sfax (ENIS), University of Sfax, Tunisia ' National School of Computer Science, University of Manouba, Tunisia ' Advanced Technologies for Environment and Smart Cities (ATES Unit), Faculty of Sciences, University of Sfax, Tunisia ' National Engineering School of Sfax (ENIS), University of Sfax, Tunisia

Abstract: Detecting text in scenes poses a significant and challenging problem due to the complex character shapes and the diverse degradations present in natural images. It represents the initial and crucial phase that must be successfully completed before the text recognition stage. In this study, we suggest a hybrid approach to tackle this issue, leveraging the maximally stable extremal regions (MSER) algorithm, which gained significant attention in recent research. Despite its popularity, it remains very sensitive to the shape, size, scale and background noise of text characters. To tackle its limits and refine the final result, we focus on an extended MSER based method. Overall flowchart of the suggested system is divided into three steps: 1) robust masks generation to identify the text candidate regions; 2) feature extraction and selection based on VGG16 deep learning architecture; 3) employing classifiers in a cascaded structure to differentiate between text and non-text areas based on enhanced geometrical pattern characteristics. The effectiveness of the proposed method is demonstrated through an experimental study conducted on various benchmarks, such as ICDAR2013, ICDAR2015, MSRA-TD500, and RRC-MLT.

Keywords: scene text; text detection; maximally stable extremal regions; MSER; VGG16 deep learning model; random forest; SVM; NCA-based feature selection.

DOI: 10.1504/IJCSE.2025.146076

International Journal of Computational Science and Engineering, 2025 Vol.28 No.3, pp.255 - 266

Received: 26 Mar 2023
Accepted: 23 Jan 2024
Published online: 06 May 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Scene text detection using robust masks and cascaded classifiers

Keep up-to-date