Techniques For Text Extraction In Videos

To some extent, the high-level semantic information reflects the content of the video,and the text embed in videos contains a wealth of high-level semantic information. Iftext can be detected, segmented and recognized automatically, there will be a bigimprovement in high-level semantic information understanding, indexing andretrieval. Text extraction system is normally composed by four parts, namely, textevent detection, text area localization, text segmentation and character recognition.This paper concentrates on the algorithm research of text area localization andtext segmentation. For text localization, two methods are proposed. The first one isbased on wavelet transformation, which applies corner response image and combinedhigh-frequency sub band. And the statistical characteristic is extracted as the featurevector for K-means cluster. Then, some heuristics are employed to remove the falsepositives. This approach utilizes the classification of unsupervised learning method,therefore, it is able to avoid sample training, which is time-cosuming. The secondalgorithm is based on Gabor transformation, and it aims at Chinese text detection andlocalization. The strokes of most Chinese characters are oriented in four directions, soGabor with various scales and directions can describe it well. The cluster result fordifferent scales of Gabor transformation is combined to obtain the location of text.Experimental results show that the proposed method is robust even in low-contrastcondition.For text segmentation, in order to obtain a better recognition result from OCRsoftware, text segmentation part differentiates the text pixels from the backgroundpixels. A color space based segmentation approach is proposed, which utilizes theclassical OTSU algorithm and RGB space to find the initial text pixels. And thenK-means cluster is applied to find the final text pixels from the initial one. The binaryImage can describe the text pixels well. Experimental results show that our textsegmentation technique outperform other classical Threshold segmentation algorithm.