The segmentation and recognition modules are usually implemented sequentially in most traditional automatic license recognition (LPR) systems. In this work, we integrate segmentation and recognition into a Markov network, where bidirectional constraints between segmentation and recognition are exploited for LPR. In addition, both low-level structural attributes and compositional semantics of license plates are incorporated in a probabilistic way. A belief propagation (BP) algorithm is used for statistical inference that is able to separate and recognize license characters simultaneously. Experiments on Chinese license plates show that the proposed approach work well even when connected and distorted characters present.