A Robust Color-Independent Text Detection Method from Complex Videos

Video text carries meaningful contextual information and semantic clues for visual content understanding. In this paper, we propose a novel hybrid algorithm to fast detect video texts even under complex backgrounds. We first use an SVM classifier trained by our new StrOke unIt Connection (SOIC) operator to identify seed stroke units. Stroke shape distributions, instead of color or texture features, are extracted and trained in our method. Then the stroke units are tracked and extended into their surroundings to form text lines, obeying seed stroke geometric constraints. Experimental results show that our approach is color and language independent, and robust to video illuminations.