Devanagari Text Detection From Natural Scene Images

1. Introduction

Devanagari script is used in India for writing many official languages like Hindi (National Language of India), Marathi, Sindhi, Nepali, Sanskrit and Konkani (Jayadevan, Kolhe, Patil, & Pal, 2011). Official documents, instruction boards, street boards, banners etc. are generally written in regional languages. Though there is a significant improvement in printed script recognition from documents, there is lot of scope for research on regional text processing in scene images (Pal, Jaydevan, & Sharma, 2012). Camera captured scene images are complex due to dust, uneven light conditions, shadows, perspective distortions, poor quality etc. Also it is difficult to identify text with different styles, colours and sizes. Thus, it is very challenging to detect text in camera captured scene image as shown in Figure 1.

Rapid growth of multimedia and handheld systems demands more efficient techniques to process digital data. Text present in camera captured images is useful in many applications. An automatic text detection and recognition system can be used in various applications like content-based image searching, automatic navigation system, object recognition, text to audio conversion and language translation. Smart phone-based applications can be developed to translate information written in Devanagari script (Hindi, Marathi, Nepali, Konkani etc.) present in scene images to other target languages. Such a system will be very useful to foreigners and others who can’t read the language (script). Text detection and recognition are two major steps in such applications.

Figure 1.

Key challenges in text detection: (a) Complex background; (b) Poor quality text due to various climatic conditions; (c) Shadow effect; (d) Perspective distortion; (e) Uneven light condition; (f) Multi coloured and multi sized text; (g) Curved text; (h) Artistic font

To detect Devanagari text from scene images, two techniques are proposed in this paper. The first technique is based on edge detection and the second is based on colour information. Later the two techniques have been combined to achieve higher accuracy. Edges play an important role in multi sized, multi coloured and multi oriented text identification. In the edge-based technique, image is preprocessed before applying the edge detector. Region-based properties are used for elimination of irrelevant edges. Text regions are then identified by connected component analysis. Colour homogeneity is also an important feature for text detection in high contrast images. In the colour-based technique, background and foreground regions are separated by colour based clustering followed by region-based property filtering and connected component analysis. A dataset of natural scene images is created and experimentation is carried out to check the accuracy of individual and combined approaches.

This paper is organized as follows: Survey of existing text detection methods is summarized in section 2. The proposed techniques are described in further sections. Section 3 presents edge detection-based technique and its experimental results. Colour-based technique based on clustering in YCbCr colour space is mentioned in section 4. Finally, these two techniques are combined to achieve higher accuracy. The combined approach is explained in section 5. Observations of all these techniques are concluded in section 6.

2. Related Work

Plenty of work is carried out on English text detection and recognition from scene images. There is some improvement in research related to languages like Indic scripts (Jayadevan et al., 2011; Pal et al., 2012), Farsi/Arabic (Maryam, & Mohammad, 2012) and Chinese script (Bai, Chen, Feng, & Xu, 2014) for the last few years. The state of the art of various text detection & recognition techniques are mentioned by Zhu et al. (2016), and H. Zhang et al. (2013). Text detection is the process of identifying text region present in the image. In text recognition, the detected text from the image is converted into a machine readable format. Text present in camera captured images is different in style, size, colour and orientation. It is very difficult to apply a common method for text identification due to this variation. The work becomes more challenging due to complex background, perspective distortion, poor resolution, dust and shadow effect. Different methods that can be used for text detection are connected component, edge-based, colour-based, texture-based and stroke-based methods (H. Zhang et al., 2013; J. Zhang, Cheng, Wang, & Zhao, 2013).