Study On Offline Handwritten Chinese Signature Verification Based On Data Fusion Scheme

The main difficulty of offline handwritten Chinese signature verification is feature extraction. So how to extract features based on characteristics of Chinese signature is discussed in this dissertation. When signature verification is considered, there are two problems needed to solve. One is static feature extraction. The other is seeking dynamic features. And this is mainly concerned in this dissertation.When static features are extracted, they are described using pseudo-Zernike invariant moments calculated based on thinned signature images. Several important dynamic features are found, and a very important ratio feature is extracted using wavelet transform. In addition, when signature shape and high-density areas are depicted in structural way, singular value decomposition of matrix is utilized. As to classifier designation, first simple Euclidian distance classifier is used, then back-propagation network is utilized. At last a data fusion scheme is realized. The main works of this dissertation are as follows.1. As to pre-processing of signature image, a new thinning algorithm is proposed aimed at questions existing in normal thinning algorithms. The algorithm is simple, but branch problems can be solved by using it when turn strokes are thinned. In addition, a new 8-connected neighbors detection algorithm is presented. It can extract connected areas in signature image for next analysis. And new methods are proposed in gray level signature image and high-density area extraction.2. A novel signature verification scheme is proposed based on invariant moments and dynamic features which considering static features and dynamic features. Static features of signature are described using pseudo-Zernike invariant moments based on thinned signature image. 10 orders of modular values are computed as features using scale and translation invariance and anti-noise character of pseudo-Zernike moments. When dynamic features are extracted, global and local high-density areas are obtained from gray level image at first. Then global and local high-density factors are calculated as ratio of area of high-density image to that of signature image. In addition, based on global high-density area, the relative gravity center is computed as another feature. So 16 features are obtained and a verification system is built using Euclidian distance classifier. Experiment result shows FAR(false acceptance rate) and FRR(false rejection rate) can achieve 7.25% and 9.30%. 3. In further research of dynamic features, a novel method is proposed using the ability of describing appropriate information of wavelet transform. When wavelet transform is used, a forth discrete wavelet transform is performed on weighted normalized histogram of gray level signature image with Daubechies(4). And the forth appropriate coefficients are reconstructed. Unlike usual use of wavelet transform on extracting detail information, appropriate information is obtained. And a ratio is calculated as another important feature. Weighted Euclidian distance classifier is used based on the former research. Experiment result shows that FAR and FRR can achieve 7.83% and 6.88%4. Connected with BP neural network as classifier, a novel verification system is constructed with the features extracted as network input. Experiment result shows its efficiency.5. Based on former research, a novel data fusion scheme is proposed using singular value decomposition for describing features from signatures. And a novel hybrid scheme which is a novel data fusion scheme is proposed combing it with the former one. The final output is obtained through competitive selection of outputs of 2 BP networks. And a verification system is built which is a classifying system combining 2 classifiers. Based on thinned signature image and high-density image, 40 features composing feature vector are extracted with SVD. And they are the input of the BP network. Then the final output is produced with outputs of this network competed with those of the former network. Experiment result shows that it's more effective than any of 2 lonely systems. FAR and FRR can achieve 5.71% and 6.25% respectively.