Mobile Text Localization on the GPU

Extraction of text from camera images is a non-trivial task, especially when
performed on mobile devices. Current applications (business card readers)
often suffer from tedious handling and varying quality. 

Typically the location of text is determined by a suitable algorithm and the region ist handed over to an ocr engine which computes the final text string. However, sophisticated algorithms for text localization often have a comparatively large runtime. So, they cannot be directy applied to mobile devices.

In a first step an exemplary implementation of a text localization algorithm shall be studied, optimized (e.g. FastCV) and evaluated. In a second step, performance-critical parts shall be identified and subsequently accelerated using the cababilites of modern mobile devices. In particular, operations shall be performed on the mobile GPU (e.g. OpenCl).

We further assume that the text is located on a planar target (e.g. store-card)
that can be tracked. By exploiting the available pose information and combining
different views of the target, a more detailed text mask (possibly of higher resolution),
suitable for a subsequent ocr stage, shall be extracted. 

Requirements / Opportunities
Basic lectures in image processing / graphics
Matlab / C++
Interest in mobile hardware


References

[1] B. Epshtein, E. Ofek and Yonatan Wexler. Detecting Text in Natural Scenes with Stroke Width Transform. In CVPR, pages 2963-2970, 2010

[2] D. Wagner, G. Reitmayr, A. Mulloni, T. Drummond, and D. Schmalstieg. Pose Tracking from Natural Features on
Mobile Phones. In ISMAR, pages 125–134, Sep. 2008.

[3] D. Wagner, A. Mulloni, T. Langlotz, D. Schmalstieg. Real-time Panoramic Mapping and Tracking on Mobile Phones. In VR, pages 211-218, 2010

• AR4DOC - Augmented Reality for Document Inspection