The key of license plate recognition system is character recognition, and the core of character recognition is to extract character features. Wavelet transform can obtain the details and structure features for characters, and the invariant moment can describe it well, which are combined to extract the character features. The character strokes feature is extracted by using the directional feature of high frequency sub-images which are decomposed by tensor product wavelet, and the alliance feature vector reflecting the structural and statistical features of characters is obtained. The decomposition of character images adopts the second generation lifting wavelet algorithm, which further reduces the computational complexity. Experiment results show that the alliance feature vector extracted by the proposed method can achieve 98% character recognition rate, which can meet the requirements of practical application.