Off-line automatic signature verification techniques: Flexible template matching approaches
MetadataShow full item record
The objective of this dissertation is to discriminate the forgery off-line signatures from the genuine signatures through intelligent image analysis. In the field of off-line signature verification researchers devote mostly to design the powerful feature set and direct matching methods in either pixel or stroke level. But few of them try to combine the good aspects of both approaches. Motivated by that two new methods are developed independently but designed in the same perspective of thinking: matching with the global shape and comparison in the local correspondence. The first method matches the constructed pseudo-dynamic paths of two signature samples being compared by means of dynamic time warping. And the dissimilarity of the two signatures is measured by the harmonic distance of moments features of curve segments. Following the principle of detection and comparison in pattern recognition the second method tries to find a surface to surface mapping between two signature images in the way of mathematic morphology and spline warping. 20 types of extrema along the stroke contours are defined and extracted to construct the mapping function between the signatures being compared. GSC---the gradient, the structural and the concavity features are extracted so far from the pairs of matched regions. Methods to eliminate erroneous correspondences were developed from the perspectives of extrema types and deformation evaluation. Chaincode processing plays an important role in both of two methods. From it the first method extract the moment feature and the second one finds the local extrema. Mathematical morphology operations are designed to smooth the edges of strokes and connect the broken strokes which are introduced in either the phase of fast writing or that of scanning. We also designed a method to remove the print text content surrounding the signature image. This method is an extension to existing page decomposition models. Font size features are extracted through discrete component analysis. Fisher linear discriminant and Bayesian classification are used to generate suspected components. Consequently spatial relation of the components are referred to determine the final component set to be removed. We used error trade-off curves to evaluate the discriminative power of proposed approaches. Especially Zernike moments, wavelet moments and Fourier moments are compared in the first method. The performances of individual features are also evaluated. Additionally distance measurement, KNN, SVM and maximum likehood are applied independently to classify the forgery and genuine signatures. The experiments demonstrate the efficiency of matching phase designed for shape feature extraction. The equal accuracy rate of flexible template matching method reaches 92% which is better than previous methods and comparable to that of on-line systems.