Abstract: A new algorithm for license plate character segmentation based on differencing projection and preferably segmented characters is proposed. The level differencing projection is needed to correct skew character with level segmentation to implement simultaneously, thus reducing the complexity of the algorithm with improved positioning accuracy. Then the preferably segmented characters of a license plate can be found by combining vertical projection with connected domain algorithm to generate the sliding template for vertical segmentation of characters. In this way the problem of recognizing adherent and/or broken characters can be solved efficiently. Experimental results proved that the algorithm proposed has advantages of fast segmentation and preferably segmented characters, and self-learning ability and strong anti-jamming performance to recognize ambiguous characters on license plates.