For digital video in complex scene, a multi-scale technique was fused into the conventional gradient detection, and a two-stage coarse-to-fine text detection algorithm was proposed. The features of weighted averaging gradient energy and motion energy were exploited to coarsely detect sample frame and obtain the candidate text regions. The false detected candidate text regions were filtered by connected-component analysis. The statistical characteristics of gradient direction were further utilized to obtain the fine detection results. Experimental results have shown that the proposed approach is robust for digital video with complex backgrounds.