A Character Feature Extraction Based On Dual-tree Complex Wavelet Transform And Recognition

This paper begins with the definition of traditional real wavelet transform, and then explores the principium about how to decompose and reconstruct signal with it. The Dual-Tree Complex Wavelet (DTCWT) is a new wavelet transform. It not only keeps the multi-resolution and the analytic ability to time-frequency localizability, but also provides good directional selectivity, shift invariance and limited data redundancy which are lack in the traditional wavelet transform. It is just the shift invariance of DTCWT solved the problem, variation of energy in each level of coefficient for the shift of signal, which can not be solved using real wavelet transform; The good directional selectivity makes DTCWT get eight subbands in every level, the high-frequency which reflect detail feature can describe six directional characteristics ( ±1 50 ,±4 50 , ±7 50) respectively, so it can obtain more information about gray graph in different level. In this paper, we combine DTCWT with Singular value decomposition (SVD) to apply to car license plate Chinese feature extraction, using the former to decompose and reconstruct the gray graph which have been preprocessed in advance and get one low-frequency subgraph and six high-frequency subgraphs, then extract every subgraph's stabilized singular values by means of singular value decomposition to constitute a feature vector. We make a dimension descent and unitary processing to feature vector for the purpose of improving the computational efficiency and convergence rate before put it into BP neural net to train and simulate. The experimental results demonstrate this method can get good recognition rate and the contrast to real wavelet transform declares it is more efficient than the latter.