A new signature verification technique based on a two-stage neural network classifier

https://doi.org/10.1016/S0952-1976(00)00064-6Get rights and content

Abstract

This paper presents a new technique for off-line signature recognition and verification. The proposed system is based on global, grid and texture features. For each one of these feature sets a special two stage Perceptron OCON (one-class-one-network) classification structure has been implemented. In the first stage, the classifier combines the decision results of the neural networks and the Euclidean distance obtained using the three feature sets. The results of the first-stage classifier feed a second-stage radial base function (RBF) neural network structure, which makes the final decision. The entire system was extensively tested and yielded high recognition and verification rates.

Introduction

As available computing power eventually increases and computer algorithms become smarter, tasks that a few years ago seamed completely unfeasible, now come again to focus. This partly explains why a considerable amount of research effort is being recently devoted in designing algorithms and techniques associated with the problems like human handwritten signature recognition and verification.

A signature recognition and verification system (SRVS) is a system capable of efficiently addressing two individual but strongly related tasks: (a) identification of the signature owner, and, (b) decision whether the signature is genuine or forger. Depending on the actual needs of the problem at hand, SRVSs are often categorized in two major classes: on-line SRVSs and off-line SRVSs. While for systems belonging to the former class, only digitized signature images are needed, for systems in the latter class, information about the way the human hand creates the signature such as hand speed and pressure measurements, acquired from special peripheral units, is needed.

During the last few years, several on-line Parizeu and Plamondon, 1990; Brault and Plamondon, 1993; Lee et al., 1996) and off-line (Qi and Hunt, 1994; Yedekcoglu et al., 1995; Han and Sethi, 1996; Droughard et al., 1996; Bajaj and Chaudhury, 1997; Huang and Yan, 1997) SRVSs have been proposed. In the off-line category, Qi and Hunt (1994) proposed a SRVS that is based on global and grid features in conjunction with a simple Euclidean distance classifier. Yedekcoglu et al. (1996) developed a technique based on thickened templates that can be utilized as an initial face of a SRVS in order to reject signatures that are completely unmatched. Han and Sethi (1996) proposed a signature retrieval and identification system based on geometric and topologic features. Droughard et al. (1996) used directional probability density function in conjunction with backpropagation-trained neural networks. Bajaj and Chaudhury (1997) used multiple neural networks supplied by three sets of global features, including projection moments. Huang and Yan (1997) use geometric features in combination with a neural network classifier. However, the experimental results were based on a small number of samples. Ramesh and Murty (1999) propose a system for off-line signature verification, which consists of four subsystems based on geometric features, moment representations, envelope characteristics and wavelet features.

In this paper, a novel approach for off-line signature recognition and verification is proposed. The presented system is based on three powerful feature sets in combination with a multiple-stage neural-network-based classifier (Fig. 1). The novelty of the system lies mainly on the structure of the classifier and the way that it is used. The neural network classifier is arranged in two stages.

We have understood that for a SRVS to be functional in practical applications, the ability to easily add/remove signatures from new/obsolete owners to its database must be inherent. Our approach towards this goal is to implement the structure of the neural network classifier is a one-class-one-network scheme. That is for each signature owner an individual classifier is being implemented. Each time signatures from a new owner are added to the SRVS database, only a small, fixed-size, neural-network-based classifier must be trained.

Moreover, to farther overcome training difficulties stemming from the feature set size, the proposed feature set is divided into three individual feature groups of different physical meaning. For each of the resulting three feature groups, an individual multi-layer perceptron (MLP) neural network is implemented. These three small and fixed size neural networks for each signature owner constitute the first stage of the classifier. It is a task of the second-stage classifier, a radial basis functions (RBF) neural network to combine the results of the first stage to make the final decision of weather the presented to the system signature, belongs to a candidate owner or not.

The experimental results confirm the effectiveness of the proposed structure and show its ability to yield high recognition and verification rates.

Section snippets

Preprocessing

The preprocessing stage is divided into four different parts: noise reduction, data area cropping, width normalization and signature skeletonization.

Feature extraction

The choice of a powerful set of features is crucial in optical recognition systems. The features used must be suitable for the application and for the applied classifier. In this system, three groups of features are used categorized as global features, grid information features and texture features.

While global features provide information about specific cases concerning the structure of the signature, grid information and texture features are intended to provide overall signature appearance

The signature database

For training and testing of the SRVS many signatures are used. The results given in this paper are obtained by using a signature master database of about 2000 signatures. The signatures were taken from 115 persons (15–25 signatures from each).

For training the system, two subsets, taken from the master set, of about 1000 and 500 signatures were used. The first subset (TRS1) was used to train the first-stage classifier while the second subset (TRS2) was used to train the second-stage classifier.

Classification

Multi-layer perceptron (MLP) neural networks are among the most commonly used classifiers for pattern recognition problems. Despite their advantages, they suffer from some very serious limitations that make their use, for some problems, impossible. The first limitation is the size of the neural network. It is very difficult, for very large neural networks, to get trained. As the amount of the training data increases, this difficulty becomes a serious obstacle for the training process.

The second

The training phase

The training of the system includes the following two steps.

Step 1: Train the first-stage classifier

This task consists of training the three neural networks for each person in the TRS1. The TRS1 consists of 1000 signature images randomly selected from the master set of 2000 signature images. There are available signatures from 115 different persons and so we have to train 115×3=345 different but small size neural networks.

Each neural network corresponds to a specific owner, and therefore, all

Testing phase and results

According to the above analysis, when the system is asked to decide whether an unknown signature image belongs to a particular person in the database the following steps are followed.

  • The unknown signature image passes through the pre-processing and feature extraction stages.

  • The three sets of features are applied to the inputs of all of the three specialized Perceptron neural networks. The networks are run forward so that we get outputs for all of them.

  • The Euclidean distance between the 160

Conclusions and remarks

This paper proposes a new off-line signature verification and recognition technique. The entire system is based on 160 features grouped to three subsets and on a two-stage neural network classifier that is arranged in an-one-class-one-network scheme. During the training process of the first stage, only small, fixed-size neural networks have to be trained, while, for the second stage the training process is straightforward.

In designing the proposed system, most of our efforts were towards of

References (15)

There are more references available in the full text version of this article.

Cited by (0)

View full text