Title: A two-stage text detection approach using gradient point adjacency and deep network

Authors: Tauseef Khan; Ayatullah Faruk Mollah

Addresses: Department of Information Technology, Haldia Institute of Technology, Haldia 721657, West Bengal, India; Department of Computer Science and Engineering, Aliah University, IIA/27 New Town, Kolkata 700160, West Bengal, India ' Department of Computer Science and Engineering, Aliah University, IIA/27 New Town, Kolkata 700160, West Bengal, India

Abstract: Scene text detection is a pivotal problem in computer vision and image processing research. In this paper, a fair attempt is made to design a simple yet effective text detection method in Indic script environment. At first, a fine-scale edge-map is generated from the original image, and subsequently, adaptive clustering is applied to form clusters of edge-points based on their spatial density. Foreground objects are extracted with the help of cluster boundaries and considered as prospective text proposals. Such text proposals are fed to a deep convolutional neural network for learning and prediction as text and non-texts. Finally, true-text components are aggregated as localised final texts of the original image. The proposed method is evaluated on two benchmark datasets viz. ICDAR 2017-MLT and ICDAR 2013 born image, and obtained results are found to surpass some other state-of-the-art methods, which demonstrates its effectiveness in scene and digital environments.

Keywords: text detection; text proposal; convolutional neural network; foreground object classification; text proposal aggregation; born-digital images; scene images.

DOI: 10.1504/IJCSE.2022.122210

International Journal of Computational Science and Engineering, 2022 Vol.25 No.2, pp.152 - 165

Received: 07 Dec 2020
Accepted: 29 Apr 2021
Published online: 12 Apr 2022 *

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