In this paper we present a novel approach to detect texts in video frames. The approach proposes a spatio-temporal wavelet transform to integrate information of multiple frames rather than a single one. Static and dynamic texts are detected separately due to their characteristics in temporal domain. Sub-bands decomposed from the original image sequence are combined to form a salience map, which features are extracted from. The approach is verified by experiments with various types of videos. High average recall and precision rates confirm the effectiveness of the proposed method.