Signature verification is one of the most widely used ways to authenticate any person in today’s world. Due to this reason, signature forgery is being widely attempted by attackers. Signature verification is classified into two categories online signature verification and other is offline signature verification. This paper focuses on identifying forgery done during the online signature verification. In this paper, we used the signature verification competition 2004 (SVC2004) dataset for our experiments. In the proposed method we used Discrete Fourier transform which is used to extract distinguishable features between forged and genuine signatures. In the next step, we used Recurrent Neural Networks methods Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) for classification. We used bidirectional LSTM and bidirectional GRU as here we do have past as well as future results. Final results using bidirectional methods outperformed our overall results.

Online Signature Verification Using Bidirectional Recurrent Neural Network

Handwritten Signature verification is an important personal identification. It is widely used in authorizing a cheque or legal documents. Signature verification is either online (dynamic) or offline (static). A machine learning, Haar Cascade Classifier (HCC) approach was introduced by Viola and Jones [1] to achieve rapid object detection based on a boosted cascade of Haar-like features. Here, for the first time the HCC approach was applied for the handwritten signature recognition and verification. Two datasets were used, UTSig dataset [2] for Persian writers which written from right to left and includes 8,280 images from 115 writers. GPDS synthetic Signature database [3] for English writers which written from left to right. It contains data from 4,000 synthetic individuals. A classifier was created for each one of the writers’ signatures after applying the preprocessing phase. Each classifier was trained and tested using enormous number of signatures generated from applying artificial noises on the signature images. The system was tested with real signature images and produced accuracy of 92.42% with UTSig dataset and 92% with GPDS dataset. The proposed approach achieved high results compared to approaches built using traditional machine learning approaches. The approach proof applicability with the different types of writers’ languages and style.

Handwritten Signature Verification using Haar Cascade Classifier Approach

Signature is one of the important characteristics that can be used in verification of several documents, one of which is an academic document. Signature verification in the academic environment is important, especially in ensuring the authenticity of the lecturers or teaching staff signatures. Not a few students who choose to falsify the signature of a lecturer or teaching staff in order to facilitate their academic process, this is an important issue especially if the student is actually not eligible and does not meet the criteria to get a signature or endorsement from the lecturer concerned. A technique or method is needed that can help the process of verifying the signatures of lecturers and teaching staff in an academic environment. One technique that might be used is to use image processing techniques. In this study a classification will be made between the genuine and forgery signature images as a verification process of the authenticity of the lecturer signatures obtained by students. The data used is the signature image of a lecturer at Amikom Purwokerto University who was a examiner at the Practical Task Seminar. The method proposed in the classification process uses the Support Vector Machine (SVM) algorithm with two different kernels. Both kernels consist of poly kernel and normalized poly kernel, the selection of the two kernels is used to compare which results are more optimal. The results of this study are SVM by using poly kernel normalization to be able to give better results when compared to using poly kernel only. The results obtained using poly kernel normalization are an accuracy level of 79.43% and a specificity level of 100%.

Analisa Hasil Perbandingan Poly Kernel Dan Normalisasi Poly Kernel Pada Support Vector Machine Sebagai Metode Klasifikasi Citra Tanda Tangan

Signatures are one of the important characteristics that security needs to be considered. Some cases related to signature forgery often occur, this is certainly dangerous especially if the signature forgery can be misused. So there needs to be a verification process on the authenticity of signatures related to this. Several studies related to signature verification have been carried out, one of them using digital image processing techniques. However, some studies only propose a method without comparison of results. This study aims to compare methods and development of signature verification methods based on digital image processing with machine learning techniques. The final results of this research can later be used as a design module that can be used in system development or signature verification applications. The data used is the image of the digitization of the signature of the Lecturer in the STMIK AMIKOM Purwokerto environment. The segmentation method used in this study is adaptive maximum minimum thresholding with double morphological operation. Good segmentation results are expected to provide good classification results. Comparison of several different classifiers in the classification stage is carried out, including Linear Regression, Naive Bayes (NB), Support Vector Machine (SVM), Multilayer Perceptron (MLP) and K-Nearest Neighbor (K-NN).

Double Morphological Segmentation for Increasing Performance of Signature Classification Using Machine Learning Technique

Automatic biometric recognition for individuals is a challenging problem, which has become increasingly popular in recent years by various purposes in different fields. Face and signature identification is the main issues, and until the date, no methods deliver a powerful solution to all circumstances. In this work presents an advanced technique is used for both handwritten signature and human face recognition. Some of the most common methods used in image-based approaches are 2D-DCT, data removal from images, and self-organizing maps utilising neural networks. The discrete cosine extracts method extracts the feature in face and handwritten signature based on the image color. The construct feature vectors calculate the DCT coefficients. A neural network classifier with SOM is used to categorize feature vectors based on the DCT which are used to classify the problem is present or not in the database. Both of the biometrics images are accomplished with the power rates of the grayscale pixels into many groups, and it also trained with the neural network classifier. The obtained result was evaluated and executed in MATLAB environment using a various image database of the face and handwritten images which consist of different dimensional images. Also, it observes the variation of efficiencies of the system for multiple numbers of deep layers and epochs to make comparison and contract along with a high recognition rate of 83.06% for every five trails. This technological advantage is both memory and speed usage, which makes high-speed performance and less computational requirement. 

A Self-organizing Map with Neural Networks Classifier Technique for Face and Handwritten Signature Recognition System in DIP

Very large percentage of daily financial transactions is generally carried out on the basis of verification of signatures. Therefore signature plays an important role both for authentication and authorization of any legal documents. Signature verification is the process used to verify an individual's hand written signature is genuine or forged signature. Classification of recognition rate of genuine signature and rejection rate of forgery signature is 95% and 5% respectively.

M. K. Prathiba

Papers

Signature Verification System Based on Wavelets