Research on the rejection capabilities of the signature-pressure-based individual recognition system for counterfeit signatures using optimized neuro-template with Gaussian function
Lina Mi · Institutional Repositories DataBase (IRDB) · 2007
In recent years, internet business has been developing rapidly along with the wide application of Internet and the security of users' information has become more and more important point for the development of internet business.Therefore developing an individual recognition system especially fit for internet application is becoming pressing task.Recently, biometrics information, such as finger print, iris, voice, face and signature, has been increasingly adopted in personal identification because of being unique and having high resistance to forgery.In this research, a novel individual recognition system, in which biometrics information of signature pressure is exclusively employed to present personal feature, is developed for the application of internet business.The execution of the system includes two procedures, registration and recognition.First the user give three register signatures to register on the system (registration), after registration, the user can log on the system by giving one test signature (recognition) at anytime.In both procedures, the signature pressure data will be preprocessed firstly and then the data obtained from preprocessing of source pressure data are used either for registration or for recognition.Therefore two parts, preprocessing and neural network classifier (NN classifier), are included in the structure of the system.In the preprocessing, the detected signature pressure data is firstly normalized, then equally dividing the normalized data into 300 sections and average value of each section is calculated as element of relay data.Second, validity check is executed on three relay data of register signatures.Last, the probability distributions of register relay data and inhibit relay data, which are artificially made by system, are analyzed and 50 elements are extracted as slab value from each relay data, which are used for NN learning or input to NN.In the preprocessing, the scale of source data is greatly reduced and personal feature of signature pressure are also extracted.The neural network classifier of system is mainly studied in this paper and the uniqueness works in the research of this paper mainly include the following points.1) Neuro-template Matching Method is introduced into the NN classifier of the system.According to this method, each registrant is assigned with a three-layer feed-forward neural network with uniform structure of 50×35×2 and the NN classifier of the system is composed with the neuro-templates of all registrants.In case of registration, after learning with preprocessed source data of register signatures as samples a new neuro-template is constructed for new registrant and then adopted as part of NN classifier by the system.In case of recognition, preprocessed pressure data of the test signature is matched with each existing neuro-template, and then outputs of all neuro-template are evaluated to decide the identity of signer.The performance of our system shows that Neuro-template Matching Method successfully simplified the registration procedure and removed the limitation on the number of registrants of the system.Furthermore, according to the study on mutual influence among neuro-templates, relearning of existent neuro-template caused by recruitment of new neuro-template is individual-dependent, and helpful for rejection capabilities of relearned neuro-templates for counterfeit signatures.2) Gaussian ridge function is proposed as activation function of neurons in hidden layer and output layer of neuro-template to improve rejection capabilities of system for counterfeit signatures.Though the developed signature-pressure-based individual recognition system is effective in recognizing the authentic signatures, it suffers from poor rejection capabilities for counterfeit signatures.In order to improve rejection capabilities of the system on premise of ensuring the recognition capabilities satisfied, a kind of Gaussian function is proposed as activation function of neurons in hidden layer and output layer of each neuro-template, instead of originally employed sigmoid function.Different traditional Gaussian function of radial basis function (RBF), proposed Gaussian function is a ridge-like and semi-localized function, the corresponding sensitive field of function stretches out infinitely along the ridge and is restricted by the width parameter on the orthogonal direction of the ridge, and its sensitive field is controlled by the width parameter.The neuro-template with Gaussian ridge function presents the distribution of patterns of all known categories instead of partitioning the feature space as sigmoid function does.Hence the neuro-template with proposed Gaussian ridge function has more potential to improve rejection capabilities of the system with ensuring recognition capabilities of the system.The experiment results show that the employment of proposed Gaussian ridge function effectively improved the rejection capabilities of the system comparing with original system based on sigmoid function, at same time, however, it also led to slight decrease of the recognition capabilities of the system.3) Width parameter sigma of Gaussian ridge function is customized for each neuro-template using improved back propagation method (BP).In the pilot study.The width parameter of proposed Gaussian ridge function is selected manually and kept constant once proper value is decided.Though the neuro-template based on Gaussian ridge function with fixed sigma is effective on improving the rejection capabilities of system for the counterfeit signatures, the improvement is not significant enough.Moreover the recognition capabilities of system for genuine signatures become deteriorated a little by the employment of Gaussian ridge function with fixed sigma.To further improving the rejection performance of the system at same time ensuring recognition performance satisfied, the width parameter of Gaussian ridge function is optimized for each neuro-template.The experiment results showed that the customization of width parameter sigma not only effectively furthered the improvement of rejection capabilities of the system, but also partially recovered the deteriorated recognition capabilities of the system resulted by employment of Gaussian ridge function.4) The uniformed characters are proposed as register characters of the system.Though the performance of our system has been improved by employment of proposed Gaussian ridge function with optimal width parameter sigma, the high discrepancy in the recognition capabilities of the system for different registrants has been seen and that indicates the insufficiency of performance stability of the system.To address that problem, uniformed characters is proposed as register characters of the system, for which personal signature is traditionally employed, in the last part of this research.To evaluate the feasibility and effectiveness of proposed method, four groups of characters are selected as uniformed register characters respectively and corresponding recognition capabilities of the system are tested and compared with that of original signature-based system.The experiment results show that the uniformity of register characters seems to be effective in reducing the fluctuation of the recognition capabilities of the system for different registrants and improving the performance robustness of the system without sacrificing the recognition performance itself.Furthermore the discrepancy in the performance stability of the systems with different uniformed register characters shows that the robustness of performance of the system is influenced by the complexity of the register characters, it suggests that uniformed register characters should be carefully selected to get better performance stability of the system and better performance itself.Last, from experiment results it also can be seen that the performance robustness is almost not affected by the employment of Gaussian ridge function and optimization of parameter sigma, which are proposed to improve the rejection capabilities of the system for counterfeit signatures.