Stefania Calarasanu (1), Séverine Dubuisson (2) and Jonathan Fabrizio (1) 

                (1) LRDE-EPITA, 14-16, rue Voltaire, F-94276, Le Kremlin
      Bicêtre, France

   (2) CNRS, UMR 7222, ISIR, F-75005, Paris, France

Abstract:

A frequent challenge for many Text Understanding Systems is to 

tackle the variety of text characteristics in born-digital and natural 

scene images to which current OCRs are not well adapted. For 

example, texts in perspective are frequently present in real-word 

images, but despite the ability of some detectors to accurately 

localize such text objects, the recognition stage fails most of the 

time. Indeed, most OCRs are not designed to handle text strings 

in perspective but rather expect horizontal texts in a parallel-frontal 

plane to provide a correct transcription. In this paper,  we propose a 

rectification procedure that can correct highly distorted texts, subject 

to rotation, shearing and perspective deformations. The method is 

based on an accurate estimation of the quadrangle bounding the 

deformed text in order to compute a homography to transform this 

quadrangle (and its content) into a horizontal rectangle. 

The rectification is validated on the dataset proposed during the 

ICDAR 2015 Competition on Scene Text Rectification.