Urdu Caption Text Detection using Textural Features

Abstract

The amount of multimedia data has increased manifolds in the recent years. This calls for development of efficient retrieval techniques. Among various aspects of content based retrieval, textual content appearing in videos and images serves as a powerful semantic index. Development of such a retrieval system requires detection of text regions, recognition of detected text and generation of indices on keywords. Among these, the focus of the present study lies on detection of textual content from video frames. More specifically, we target the caption Urdu text appearing in News and entertainment channels. A series of image analysis operations is first carried out to identify candidate text blocks in the image. Features extracted from text and non-text regions using Gabor filters and Curvelet transform are fed to two classifiers namely artificial neural network and support vector machine. Evaluations on a database of 1000 video frames reported promising precision and recall.