A novel approach to text detection and extraction from videos by discriminative features and density

Text is very important to video retrieval, index, and understanding. However, its detection and extraction is challenging due to varying background, low contrast between text and non-text regions, and perspective distortion. In this paper, we propose a novel two phase approach to tackling this problem by discriminative features and edge density. The first phase firstly defines and extracts a novel feature called edge distribution entropy and then uses this feature to remove most non-text regions. The second phase employs a Support vector machine (SVM) to further distinguish real text regions from nontext ones. To generate inputs for SVM, additional three novel features are defined and extracted from each region: a foreground pixel distribution entropy, skeleton/size ratio, and edge density. After text regions have been detected, texts are extracted from such regions that are surrounded by sufficient edge pixels. A comparative study using two publicly accessible datasets shows that the proposed method significantly outperforms the selected four state of the art ones for accurate text detection and extraction.