Elsevier

Pattern Recognition

Volume 34, Issue 1, January 2001, Pages 47-61
Pattern Recognition

Fusion of correlated decisions for writer verification

https://doi.org/10.1016/S0031-3203(99)00210-1Get rights and content

Abstract

A fusion approach is proposed for improving the efficiency of writer verification systems. A short handwritten sentence is employed for this purpose. Each word of the sentence is used to tackle an individual verification problem. Then, the word-level (local) decisions are fused in order to obtain a more reliable global decision by means of the Neyman–Pearson approach. The correlation of the local decisions is extensively studied and incorporated in the fusion procedure by means of the Bahadur–Lazarsfeld expansion series. A database containing 4800 sentences is employed to validate the performance of the method. The improvement in verification performance is due to both the fusion procedure applied and the full discretion of the writer to choose his own secret word.

Introduction

In financial transactions or other high security contacts, our identity is required in order to gain access to a number of facilities. A substantial number of person verification systems incorporates handwriting which is a behavioral biometric [1], [2]. The acquisition of written patterns constitutes a non-invasive process with minimum effects on health or private rights [2], [3]. Security systems based on handwriting can be categorized according to Fig. 1. Writer verification can be carried out using either signature or handwritten words [4], [5]. Fusion of the information comprised in handwritten patterns for improving verification performance is proposed in this work.

A great deal of work has been reported in the literature for writer verification by means of signature analysis [3], [4], [5]. Signature analysis and verification is usually carried out by modeling the genuine and false samples using an adequate shape descriptor [6], [7], [8], [9], [10]. The verification efficiency of these systems is measured by means of type I (miss) and type II (false alarm) errors. Despite the fact that signature is considered an attribute which uniquely characterizes a person, it may lead to a number of drawbacks when it is employed in verification procedures. Handwriting variability as well as the capability of forgers to produce good quality specimens, diminish the reliability of the verification system. Accordingly, other handwritten patterns could be used on a complementary basis to enhance the overall system efficiency.

In this paper a method for improving the reliability of an automated handwritten signature verification system (AHSVS) using a short sentence is proposed. The employment of handwritten words is justified by the fact that handwritten text contains stable and significant information for the handwriting of a writer [5]. Additionally, the use of a sentence drawn up by the writer himself will further increase the reliability of the verification system. However, the number of words in the sentence must be small in order to avoid mistakes in memorizing its content. In our experiments a five-word sentence is used for writer verification. For this purpose, a large database was created employing 20 persons to record two different types of sentences, containing a total of 24,000 words. In the verification procedure a computationally simple feature was selected, since we are not interested in the meaning of the words but only in the general characteristics of the curves involved. Accordingly, each word is represented in the feature space by means of a granulometric feature [9]. This feature is based on morphologically processing the projective profiles of the words. For improved discrimination performance other features [4], [5], [6], [7], [8], [9], [10] can also be used for word-level decision making.

The extracted features are used to tackle an individual verification problem for each word. Thus, five decisions are obtained from each sentence concerning the identity of a specific person (word-level decision). Each decision is obtained using single hypothesis testing with weighted distance measures. These five individual decisions are combined by means of a decision fusion algorithm (DFA) so as to obtain the final and more reliable decision [11], [12]. The degree of correlation among the decisions is a critical parameter for consideration when addressing the DFA [13], [14], [15], [16], [17]. The decisions obtained from the words were found to be correlated. This is due to the fact that they resulted from words written by the same person, containing similar line attributes and in some cases, the same letters [18]. The Neyman–Pearson formulation is applied in the DFA since it is regarded as the optimal scheme [13], [17], compared to the Bayesian approach [21]. An existing procedure [17] is employed to evaluate the efficiency of the N–P test, when the second-order correlation coefficients are indexed by a simple parameter. This single parameter, called the correlation index, was found 0.15 for the decisions obtained from our database. Experimental results display a discrimination error smaller than 1% for a five-word sentence. This error can be considered satisfactory since both the probability of false alarm and the probability of detection for the word level decisions were poor (0.1 and 0.9, respectively). Improvement in the DFA performance can be achieved using more discriminative features to enhance the quality of the word-level decisions, as well as, an increased number of words in the sentence employed. Simulation results describe the dependence of this error on the number of words in the sentence as well as the correlation among the decisions.

This work is organized in the following way. Section 2 presents an overview of the proposed verification system structure and the employed database. Section 3 gives a brief description of the feature used and addresses the word-level decision procedure. In Section 4 analysis of the Neyman–Pearson decision fusion rule is provided. In Section 5 a summary of the proposed method is presented using two flow chart diagrams. A model which describes the way the correlation between the decisions affects the performance of the fusion procedure is described in Section 6. Experimental and simulation results are given in Section 7, while the conclusions are drawn in Section 8.

Section snippets

The database

The proposed system architecture for increased reliability in writer verification is shown in Fig. 2. Each person uses a specific PIN number as an index in order to enter the database in which his/her personal handwriting information has been recorded. Then, the writer is requested to write down a short five-word secret sentence along with his signature. Each word is preprocessed and features are extracted in order to produce the local binary decision ui. The set of local decisions ui is then

Feature extraction

Among the various shape descriptors that have been used for handwritten pattern representation and signature analysis are granulometrics [6]. A granulometric feature vector is employed in this work for word representation [9]. It contains spatial information about the orientation of the line segments in a handwritten pattern. Accordingly, the binary image of each word is partitioned into sub-blocks. The partition W(n,m) of the word is defined as the division of the original image into a grid of

Fusion of decisions

The decisions ui cannot be considered independent since the information content of each word is not totally different from that of the others. This is because the same type of curves, letters or even syllables are common in the words of the sentence. The classification algorithm which maps the feature space to the decision space, conveys the correlation from the vectors to the individual decisions ui. The vectorU=[u1,u2,u3,u4,u5]of the correlated decisions ui, is used in the decision fusion

Outline of the proposed method

The writer verification procedure described in the previous sections is revisited here by means of two flow chart diagrams. In this way, the steps required to train the system and thus, verify the presence of a specific writer are clearly stated. Firstly, the system is to be updated with the data corresponding to a new writer. Accordingly, the training stage described in Flowchart 1 (Fig. 7), starts with the data recording. In the first step the specific writer records his/her own secret

Correlation impact on system performance

, , implies that the parameters required to design the randomized N–P test, are the local operating points Pfai and Pmi along with the set of correlation coefficients {γ}. According to the discussion in Section 3, the probabilities Pfai and Pmi are determined from the discriminative behaviour of the corresponding feature vector. Estimation of the correlation coefficients {γ} was experimentally carried out using Eq. (26) and the available local decisions ui. Experimentation, using the described

Experimental results

The performance of the decision fusion algorithm (DFA) was tested experimentally using two different ways. Firstly, our database was employed using separately the Greek and the English sentences. The obtained local decisions accompanied by the corresponding Pfa and Pd are used as inputs to the DFA. According to the second way, the correlated decisions ui are derived using a simulation approach. This second approach, gives the possibility to accurately test the DFA since the number of the

Conclusions

Using the proposed decision fusion method, security systems based on handwritten signatures can gain further reliability in writer verification. This is achieved by means of a short handwritten sentence. The words of the sentence are used separately to derive decisions about the authenticity of the writer, and then fused for achieving higher verification performance.

Three different factors can affect the verification performance of the proposed fusion procedure. The first important design

Summary

Fusion of the information comprised in handwritten patterns for improving verification performance is proposed in this work. A substantial number of person verification systems incorporates handwriting which is a behavioral biometric. A great deal of work for writer verification by means of signature analysis has been reported in the literature so far. Signature analysis and verification is usually carried out by modeling the genuine and false samples using an adequate shape descriptor. Despite

References (21)

  • R. Plamondon et al.

    Automatic signature verification and writer identification–the state of the art

    Pattern Recognition

    (1989)
  • J.P. Drouhard et al.

    A neural approach to off-line signature verification using directional PDF

    Pattern Recognition

    (1996)
  • Y. Qi et al.

    Signature verification using global and grid features

    Pattern Recognition

    (1994)
  • B. Carter, Biometric technologies, what they are and how they work, Proceedings of CTST’ 97, Orlando, FL,...
  • J. Woodward, Biometrics: privacy's foe or privacy's friend? IEEE Proc. 85 (9) (1997)...
  • M.C. Fairhurst, Signature verification revisited: promoting practical exploitation of biometrics technology, Electron....
  • F. Leclerc et al.

    Automatic signature verification-the state of the art 1989-1993

    Int. J. Pattern Rec. Artif. Intell.

    (1993)
  • R. Sabourin et al.

    Off-line signature verification by local granulometric size distributions

    IEEE Trans. Pattern Anal. Mach. Intell.

    (1997)
  • M. Ammar et al.

    Off-line preprocessing and verification of signatures

    Int. J. Pattern. Rec. Artif. Intell.

    (1993)
  • E.N. Zois et al.

    Contour features using morphological transforms

    Electron. Lett.

    (1998)
There are more references available in the full text version of this article.

Cited by (14)

  • Biometric recognition using online uppercase handwritten text

    2012, Pattern Recognition
    Citation Excerpt :

    The morphologically processed horizontal profiles of the word are used as features. Experimenting with a database consisting of 45 repetitions of the aforementioned word written by 50 users, the proposed method achieves an identification rate higher than 96% and a verification error smaller than 2.5% In [18] the same authors propose a method based in the fusion of word-level measures taken from the different words of a short sentence. The performance of their method is tested with a database containing 4800 sentences from 20 different writers.

  • Writer verification based on a single handwriting word samples

    2016, Eurasip Journal on Image and Video Processing
  • Text-independent writer recognition using modified texture and microstructure features

    2014, PIC 2014 - Proceedings of 2014 IEEE International Conference on Progress in Informatics and Computing
View all citing articles on Scopus
View full text