Semantic keyword extraction via adaptive text binarization of unstructured unsourced video