Piece-wise linearity based method for text frame classification in video

The aim of text frame classification technique is to label a video frame as text or non-text before text detection and recognition. It is an essential step prior to text detection because text detection methods assume the input to be a text frame. Consequently, when a non-text frame is subjected to text detection, the precision of the text detection method decreases because of false positives. In this paper a new text frame classification approach based on component linearity is proposed. The method firstly obtains probable text clusters from the gradient values of the RGB images of an input video frame. The Sobel edges corresponding to the text cluster are then extracted and used for further processing. Next, the method proposes to eliminate false text components before undertaking a linearity check where the linearity of the text components is determined using their centroids in a piece-wise manner. If the components in a frame satisfy the defined linearity condition, then the frame is considered as a text frame; otherwise it is considered as a non-text frame. The proposed method has been tested on standard text and non-text datasets of different orientations to demonstrate that it is independent of orientation. A comparative study with the existing method shows that the proposed method is superior in terms of classification rate and processing time.