Robust reading, or else the ability of machines to read text in unconstrained settings such as scene images and videos, has been an important application field of document image analysis for the past 20 years. This lecture will give an introduction on robust reading systems, with a special focus on the detection and segmentation of text in real scene images.

We will review classical pipelines for text detection based on sliding window and patch based text/non-text classification, and will study the evolution of these pipelines towards selective search approaches. The nature of text will be explored and we will demonstrate the importance of context in text extraction, and the tendencies stemming from perceptual organisation. The current OpenCV implementation of text extraction algorithms will be reviewed as an open, state of the art methodology that can be used as the basis for further research.

Finally, the lecture will give an overview of the last five years of Robust Reading Competition activity, reviewing open tools and methodologies for performance evaluation. Lessons to be learnt after receiving over 2500 submissions to the Robust Reading Competition over the past year will be outlined and an outlook to the future of robust reading will be offered.
 
CONTENTS (preliminary):
·         Sliding window approaches for text detection
·         Patch based text/ no text classification, and limitations
·         Text segmentation / segmentation for recognition
·         Selective search approaches for combined text detection and segmentation
·         Open source state of the art resources (OpenCV implementation)
·         The Robust Reading Competition – overview and lessons learnt
·         Performance evaluation protocols for text detection, segmentation, word recognition
·         Next steps in robust reading: End to end systems (full recognition or word spotting), from focused text to accidental text