Text-frame classification

Text-frame classification for video is important prior to text detection and recognition to ensure that the text detection and recognition process is being conducted only for frames that consist of text. Otherwise, detecting and recognizing text on non-text frames will lead to false positive where non-text objects may be mistakenly classified as text. In this paper, we investigate the use of texture features for classifying the frame as either consists of text or not. This is because text has a different texture compared to the non-text. An experimental evaluation on Histogram of Oriented Gradient (HOG) and Scale-Invariant Feature Transform (SIFT) with Support Vector Machine (SVM), K-Nearest Neighbor (KNN) and AdaBoost classifiers has been conducted on ICDAR2003 dataset. The result indicates that SIFT outperforms the existing features for text-frame classification and Adaboost is a better text-frame classifier compared to KNN and SVM.