• PDF / 477,519 Bytes
• 9 Pages / 439.37 x 666.142 pts Page_size
• 24 Downloads / 259 Views

DOWNLOAD

REPORT

---

Interactive Scene Text Detection on Mobile Devices Jinlong Hu, Baihua Xiao, Chunheng Wang, Cunzhao Shi and Song Gao

Abstract With the increasing resolution and availability of digital cameras, text detection in natural scene images receives a growing attention. When taking pictures using a mobile device, people generally only concerned with interesting texts instead of all of the text in the image. In this paper, we propose an interactive method to detect and extract interesting text in natural scene images. We first draw a line to label a region which contains the texts we want to detect. Then a coarseto-fine strategy is adopted to detect texts in this label region. For coarse detection, we apply Canny edge detection and connected component (CC)-based approach to extract coarse region from the label region. For fine detection, some heuristic rules are specially designed to eliminate some non-text CCs and then to merge the remaining CCs in the coarse region. To better evaluate our algorithm, we collect a new dataset, which includes various texts in diverse real-world scenarios. Experimental results on the proposed dataset demonstrate very promising performance on detecting text in complex natural scenes. Keywords Text detection analysis

 Interactive  Coarse-to-fine  Connected component

J. Hu  B. Xiao (&)  C. Wang  C. Shi  S. Gao The State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China e-mail: [email protected] J. Hu e-mail: [email protected] C. Wang e-mail: [email protected] C. Shi e-mail: [email protected] S. Gao e-mail: [email protected]

A. A. Farag et al. (eds.), Proceedings of the 3rd International Conference on Multimedia Technology (ICMT 2013), Lecture Notes in Electrical Engineering 278, DOI: 10.1007/978-3-642-41407-7_28,  Springer-Verlag Berlin Heidelberg 2014

287

288

J. Hu et al.

28.1 Introduction Text detecting in natural scene images plays a very important role in content-based image analysis. However, this is a challenging task due to the wide variety of text appearances, such as variations in font and style, geometric and photometric distortions, partial occlusions, and different lighting conditions. Text detection has been considered in many recent studies and numerous methods are reported in the literature [1–6]. Most of the existing methods of text detection could be roughly classified into two categories: region-based and CC-based. Region-based methods need to scan the image at multiple scales and use a text/non-text classifier to find the potential text regions. Chen et al. [1] proposed a fast text detector base on a cascade AdaBoost classifier. As opposed to region-based method, CC-based methods first use various approaches such as edge detection, color clustering or stroke width transform to get the CCs, and heuristic rules or classifiers are used to remove non-text CCs. Pan et al. [7] adopted regionbased classifier to get the initial CCs and use the CRF to filter non-text components. M