There are still some shortcomings in the existing traditional location algorithm, so we focus on the improvement of the algorithm through the unique characteristics of video images, like edge, gradient, etc. First, we can determine which are the caption frames according to the region gradient value, then use the method of re-projection to locate the subtitles region on the basis of caption frames. The judgment of the caption frames of the video effectively reduces the complexity of the text recognition. The method of re-projection can avoid the false text region between the two lines of subtitles regions. It suits each kind of video images. The experiment indicates this method could locate the video image subtitles region well. Comparing with the other methods, it had great improving and provided a convenience for the next recognition.