Moment Invariants Based On Wavelet Transform And Its Application In License Plate Recognition

The wavelet analysis theory which has developed rapidly in the last thirty years have rela-tively mature theoretical framework and quite a wide range of applications.In the field of co mputer vision and pattern recognition,wavelet multi-resolution analysis can replace the previ-ous multi-scale Pyramid analysis to carry out feature extraction and recognition for the target objects in multi-scale.Furthermore,wavelet analysis have better application effect becase of its decorrelation in wavelet decomposition.It's a key problem to search for wavelet invariants to the transformation of scale,translation and rotation in the wavelet multi-scale recognition.Considering that the image invariant moments possess these invariant properties,this paper uses invariant moment theory to characterize this kind of wavelet invariant,explore its mathe-matical form and apply it to the license plate character recognition.While having fully understood the properties and significances of the moments theory,the vanishing moment of wavelet function and scaling function in wavelet analysis theory are deeply analyzed and studied in this paper.Then the moment invariants based on wavelet transform have been obtained with moment function theory the conclusion relating to the property of vanishing moment.Specifically,from the smoothness requirement of 1D wavelet lowpass filterH(?),this paper deduces the conclusions about the vanishing moment and ate-nuation of wavelet function and scaling function.And then,the biorthogonal wavelet moment invariants are duduced from above conclusions and a theorem relating to vanishing moment of biorthogonal wavelets.Furthermore,it has been proved that the order of the deduced wave-let moment invariants can be higher than the smooth order of wavelet function based on the analysis of the the relationship between wavelet moment invariants and smooth order.The conclusion from above quantitative analysis also suggests that the wavelet moment invariants of the longer signal have better calculation precision.The correctness of the deductions and analyzations has been verified through experiments.Due to the need of the 2D image signal applications,this paper extends above 1D wavelet moment invariant to 2D case and studies the wavelet moment invariant respectively in 2D tensor product and 2D nontensor product.On one hand.this paper obtains the mathematical expression of 2D tensor product biorthogo-nal wavelet moment invariant by using the method of tensor product.On other hand,combin-ing with mathematical form of the 2D nonseparable lowpass filter H(u,v)and its mathemati-cal conditions it must satisfy,this paper studies the smoothness of the 2D nonseparable wave-let function and scaling function wich implies from the smoothness of H(u,v).At the same time,several experiments are also provided to confirm the theoretical observations including the experiment to verify the anti-noise performance.Therefore,the comparatively complete theoretical and experimental results has been obtained.After obtaining the results of the wavelet moment invariants above,this paper proposes a novel method of license plate character recognition based on wavelet moment invariants with lifting scheme.Firstly,the license plate character image is decomposed into a low-frequency subimage and three high-frequency subimage by using tensor product wavelet.Then,the featu-re vectors of low frequency sub-image are constructed based on character approximate coeff-icient with the wavelet moments and Zernike moments.Next,three high-frequency subimages are fused to one subimage.And the stroke feature vector which reflect the detail information of the character can be constructed with the pane method proposed in this paper.Finally we combine the two kinds of feature vectors and get the alliance feature vector which can reflect the structural and statistical features of the character.At the same time,we apply the algorithm of the second generation lifting wavelet to further reduce the computational complexity.The experiments results show that the proposed method can achieve 98%recognition rate and can meet the requirements of practical application.The application for this license plate character recognition also further validates the good practical value of the wavelet moment invariant theory and lays the foundation for the further expansion of its application in computer vision and pattern recognition.