Text and non-text recognition using modified HOG descriptor
Ankit Kumar Sah, Showmik Bhowmik, Samir Malakar, Ram Sarkar, Ergina Kavallieratou, Nikos Vasilopoulos · 2017
In order to convert a document image in its editable version, an OCR engine must identify and separate the nontext regions from text regions of a given document image. In the present work, a technique is developed to classify various text and non-text regions present in a document image. For that purpose, a modified version of Histogram of Oriented Gradient (HOG) is used as a feature descriptor. Multi-Layer Perceptron (MLP) is chosen from a pool of classifier by comparing the recognition accuracy of it with two other well-known classifiers viz., Random Forest (RF), Nave Bayes (NB). The designed technique is evaluated on a dataset, containing 862 images of manually extracted regions from two standard databases namely, RDCL2015 dataset and Media Team Document dataset. The proposed system has achieved 96.44% recognition accuracy and outperformed some of the state-of-the-art feature descriptors, which have been used in the literature for the same purpose.