Article Preview
Top1. Introduction
The fundamental goal of scene text detection problem (Yao et al. 2013) is to determine whether or not there is any text in a given image, and if present, to localize and recognize it. These stages involved in developing an “end-to-end” system, are equally important for any scene text processing applications. The problem of text detection and recognition (Yao et al. 2014) from images have received significant attention in recent years due to growing demand of a number of applications.
Applications for text detection in images can be broadly categorized as visual-based, multimedia and content-based retrieval. Text in web and document images is relevant to content of the same. Street hoardings and sign-boards generally have description of place, time and doer of events taking place thus carrying important information. Automatic text recognition and translation systems help people to break language constraints. Recognizing texts on packages, containers, houses, maps, etc. has broad visual based applications. However, text detection in natural scenes is extremely difficult and challenging due to the following major factors:
- •
Wide Variation of Scenes Containing Text: In comparison to document images having dark texts on light background with regular font and standard formatting, texts in a natural scene image mostly have high diversity in fonts, styles, sizes, colors and orientations.
- •
Background Complexity: The backgrounds in natural scene images can be very complex due to heterogeneity in colors and indistinguishable foreground-background due to shade closeness. Elements like signs, fences, bricks, grasses, tree leaves and branches may give impression of true texts, thereby causing confusions and errors. This makes the task difficult for a general Optical Character Recognition (OCR) engine with raw forms of such images.
- •
Interference Factors: Various interference factors like noise, blur, distortion, low resolution, non-uniform illumination and partial occlusion, may prove to be challenging issues during the detection and recognition tasks for the text present in the image. Camera captured applications face a lot of these issues.
There are two main approaches for developing the solution for scene-text recognition (Yao et al. 2014): stepwise and integrated. In stepwise methodologies, the recognition and detection tasks are separate. They employ a framework that divides the task into segments which take care of the detection, segmentation and recognition modules one after another. Whereas in integrated methodologies, word recognition is done by sharing information between the detection and recognition tasks through optimization techniques and character classifiers. Key difference is that recognition is the main aim of the task and the strategies used in integrated methods mostly deal with that aspect. Present study deals with images containing multi-lingual texts, so a single OCR engine will not suffice. Hence, a stepwise method is followed that mainly aims to localize text regions and treat recognition as a separate module for later stage. Keeping this fact in mind, a text localization system is developed in the current work.
Text detection is the precursor to the recognition stage in the pipeline and errors only get cascaded. The localization step coarsely groups the components into possible text containing regions. These candidate regions undergo another classification to separate out some of the non-text regions during verification. The guiding principle is that the text regions display a similar underlying structure which can be represented using feature engineering to identify this invariant pattern. A bottom-up approach is applied, where the localization of text regions is performed before the application of specially trained classifier to verify those segments. One of the major advantages stems from the fact that the computation cost of the later stages through OCR reduces drastically when most of the background gets filtered in the coarse localization step. Given language independent features of multilingual OCR modules, the developed system can process texts in different languages.