Text Detection in Natural Scene Images Using a Cascade of Boosted Ensembles

SHEHZAD MUHAMMAD HANIF; LIONEL PREVOST; PABLO NEGRI

Francia

Congreso; Colloque International Francophone sur l'Ecrit et le Document; 2008

Université de Rouen

In this paper, we present a text detection and localization method. Our detection technique is based on an attentional cascade and localizer uses standard image processing techniques. Each stage in the cascade is a strong classifier trained by AdaBoost algorithm.We propose a smallset of features (39 in total) capable of detecting various type of text in grey level natural scene images. The feature set is composed of three types of block-based features : mean, standarddeviation and histogram of oriented gradients (HoG). Two weak learners, linear discriminant function and log likelihood-ratio test under gaussian assumption, are evaluated. The set of classifiers is composed of single features and combination of features. The proposed scheme is evaluated on 250 images taken from ICDAR 2003 robust reading and text locating database. The text detection rate is 86% and false alarms rate is 3x10−3 and the detector can processan image of 640x480 pixels in less than 2 seconds. The results are encouraging and detector can process an image of 640x480 pixels in less than 2 seconds.