Abstract:
State-of-the-art scene text recognition systems perform satisfactorily on samples of
benchmark datasets as long as the quality of the text in an image sample is not a ected
signi cantly by certain distortions such as blurring etc.
However, their performance may drop sharply whenever the input text appears well
outside the focus of the image capturing device or it is su ered by motion blur etc.
In this study, we considered incidental scene texts which usually exhibit much more
diversity, variability and complexity together with the common challenges of scene text
recognition compared to their counterparts which are captured by properly positioning
the camera and making possible adjustments of various image capturing parameters.
In this work, we introduce a trainable deep network that implements a super-resolution
technique as the preprocessing module on low quality scene images to boost text recog-
nition accuracy of the existing models.
There are various super resolution techniques for image available in the literature which
mainly focus on reconstructing the detailed texture of image but fails to improve the
quality of texts appearing in the image and thus the results of their recognition does
not get improved.
Here, we propose a novel text-content aware super-resolution network to improve the
quality of texts appearing in natural scene image leading to their more accurate recog-
nition by automatic methods.
Simulation results of the proposed model on the ICDAR 2015 Incidental Scene Text
dataset demonstrate its e ectiveness as an e cient preprocessing model.
Code developed as a part of this dissertation is available at: https://github.com/
AnjanGiri/Thesis.
3
