An improved technique to detect text from scene videos

摘要

It is required to detect text from natural scene videos and images because text carries an essential information. The information such as vehicle license plate numbers and traffic signs are useful in real time traffic surveillance. The information such as shop names, street names, grocery signs, product labels, advertisements, ATM instructions and so on are useful to assist blind people. Numerous text detection techniques are available. Some of them either do not detect the text or partially detect the text when text is excessively small or large. Several techniques do not detect text accurately due to higher blur or changes in illumination. Moreover, the time taken by existing techniques to detect the text can be reduced and accuracy can be increased further. In this paper, we first present the evaluation of the existing scene text detection techniques based on our identified parameters. Then we present an improved text detection technique to detect text from scene video. The proposed technique increases the text detection accuracy and decreases the computational time compared to Edge-enhanced MSER based text detection technique, especially when text is of different font size and video frames are blur.