Title: Scene text detection method research based on maximally stable extremal regions

Authors: Lei Xu; Yi Liu; Lianming Mou

Addresses: Data Recovery Key Lab of Sichuan Province, Neijiang Normal University, Neijiang, 641000, China ' Data Recovery Key Lab of Sichuan Province, Neijiang Normal University, Neijiang, 641000, China ' Data Recovery Key Lab of Sichuan Province, Neijiang Normal University, Neijiang, 641000, China

Abstract: Text information is an important basis for people to understand the natural scene image. At first, an edge-enhanced maximally stable extremal regions (MSER) text detection method based on weighted guided filtering and histograms of oriented gradients (HOG) features is proposed. Then, a two layers candidate text validation method from coarse to fine is proposed. In the first layer, a heuristic rule for validating candidate character regions is designed based on the shape features of text regions. In the second layer, the recognition of character regions is realised by using support vector machines (SVMs) with 9-dimensional features such as Hu moment invariants and stroke width transformation. The proposed method is validated by the benchmark datasets ICDAR 2013. The experimental results show that the method is comparable with other most advanced methods.

Keywords: text detection; MSER; maximally stable extremal regions; edge enhancement; SVM; support vector machine; HOG; histograms of oriented gradients.

DOI: 10.1504/IJCSM.2022.123998

International Journal of Computing Science and Mathematics, 2022 Vol.15 No.2, pp.142 - 154

Received: 21 Oct 2019
Accepted: 21 May 2020
Published online: 07 Jul 2022 *

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