Locating uniform-colored text in video frames
Vladimir Y. Mariano, Rangachar Kasturi · 2002
In this paper a method is proposed for locating horizontal, uniform-colored text in video frames. It was observed that when a row of pixels across such a text region is clustered in perceptually uniform L*a*b* color space, the pixels of one of these clusters would belong to the text strokes. These pixels would appeal as a line of short streaks on the row since a typical text region has many vertical and diagonal strokes. The proposed method examines every third row of the the image and checks whether this row passes through a horizontal text region. For a given row R, the pixels of R are hierarchically clustered in L*a*b* space and each cluster is tested whether similar-colored pixels in R's vicinity are possibly part of a text region. Candidate text blocks are marked by heuristics using information about the cluster's line of shell streaks. The detected text blocks are fused with the text regions. The method was tested on key frames of several video sequences and was able to locate a wide variety of text.