Morphological waveform coding for writer identification
Introduction
Handwritten patterns constitute the behavioral part of biometrics approach towards person verification which is not invasive in contrast to that of physiological biometrics (fingerprints or iris characteristics). Off-line writer verification systems based on signatures have been studied extensively in the past [1]. A writer verification system based on handwritten text is expected to provide discrimination results equivalent to those obtained from signatures, since text has been reported to comprise rich and stable information [2]. Furthermore, a handwritten sentence can be determined and changed by the writer at will. In high security data systems like those involved in financial transactions, the first step towards reaching a specific person's data is usually carried out by means of the personal identification number (PIN) number. However, handwritten patterns such as the signature or a word can be used on a complementary basis to improve system reliability. In order to increase further the reliability of the verification system, many handwritten words can be used by means of fusion techniques [3].
In general, writer discrimination and verification approaches based on handwritten text are hardly found in the literature [4], [5]. Security reasons or specific law restrictions have prevented serious results of significant importance on the topic from publicity [6]. To the knowledge of the authors few publications are related to writer discrimination and especially to feature extraction [7], [8]. Feature extraction from handwritten text can be carried out using approaches that resemble those of signature verification. However, features which contain information of the trace of each word are usually preferable.
In this work a writer identification method is proposed, which is based on the use of a single word. The image of the word is properly preprocessed and projected onto the horizontal direction. Projection functions have been used widely in the literature for contour feature extraction [9], [10], signature analysis [11], [12] and recognition of handwritten characters (Latin, Chinese, etc.) and numerals [13], [14]. A projection is a global shape descriptor which provides a kind of line image coding [10]. The obtained projections are segmented in parts which are morphologically [15], [16] processed in order to obtain the required feature vector. The morphological processing is a type of granulometry, i.e. the measure of area reduction through successive openings [16], [17]. Two different types of windows are applied on the segments of the projection functions to control the flow of information from one part to the other. The blanks between the letters are also considered in the formation of the feature vector.
Both the statistical properties of the feature space and the capability of the specific features for writer identification are extensively studied. This study includes the underlying pdf for the feature vector components as well as the separability of the clusters in the feature space. For this purpose a cluster separability measure is proposed and analyzed. Next, two different classification schemes are tested. Namely, the Bayesian classifier and the neural networks. In the classification procedure, the binary decision problem (writer verification: is this person he who claims to be?) and the general classification problem (writer identification: identify a writer among many others) are studied. A writer verification error smaller than 5% is achieved. The error becomes smaller while increasing feature dimensionality. A database [18] was created employing 50 writers, while an English and the equivalent Greek word were used to demonstrate that the method is language independent.
The paper is organized as follows. In Section 2 the developed database is presented. In Section 3 the procedure used for feature extraction is analyzed. In Section 4 the formation of the feature space is explained and criteria for measuring cluster intra-distance and inter-distance are presented. An extensive study on the statistics of the feature space is also carried out. Section 5 deals with the experimental performance of two classification schemes in the multicategory and the two-category case. The conclusions are drawn in Section 6.
Section snippets
The database and data preprocessing
Data acquisition and preprocessing constitute an essential step towards feature extraction and writer discrimination. Specifically, the acquisition stage affects the quality of the image, which in turn determines the reliability of the feature vector and the recognition procedure. The off-line procedures dealt in this work give full discretion to obtain good quality images.
Feature extraction
The feature extraction procedure is described in this section. The proposed feature vector is obtained by means of morphologically transforming the projection functions of the thinned images. The length of the projections is firstly normalized. Afterwards, morphological openings are applied to the segments of the projection for feature extraction.
Feature space statistics and properties
The statistical characteristics of the derived four types of features are exploited in this section and conclusions are drawn about their classification capabilities. The extent of the clusters into the feature space is examined by means of the eigenvalues of the cluster covariance matrices. Information on the correlation of the features can also be obtained from these covariance matrices. Next, the pdf of the features is exploited using the K–S fit test. The Gaussian pdf is found to be a good
Classification approaches and discrimination results
In order to evaluate the performance of the proposed features for writer discrimination, a comparative study is carried out by means of two well-established classification schemes. The conventional Bayesian approach using weighted distance measures is examined first. The simple multilayer perceptron is tested, next.
In biometrics the most common issues concerning the effectiveness of features, which potentially describe the behavior of a person, are identification and verification. In the first
Conclusions and discussion
A new feature vector is proposed for writer discrimination by means of morphologically transforming the projection function of a word. This waveform is a description of the way the pixels of the word are distributed along the direction of projection. The feature vector is formed ignoring the blanks between the letters since in this case the separability of the clusters is better. Furthermore, the use of trapezoidal windows (segments) for the formation of the feature vector results in Gaussian
Summary
In this work a writer identification method is proposed, which is based on the use of a single word. A new feature vector is employed by means of morphologically transforming the projection function of the word. First, the image of the word is properly preprocessed (thresholded thinned) and then projected onto the horizontal direction. The obtained projections are segmented in parts which are morphologically processed in order to obtain the required feature vector. The morphological processing
About the Author—ELIAS ZOIS was born in Athens, Greece in 1971. In 1994 he received his B.Sc. degree in Physics from Physics Department, University of Patras, Greece. In 1997 he received his M.Sc. in Electronics from Electronics Laboratory, University of Patras, Greece. His Master Thesis was on handwritten text analysis and writer identification. He is now a research fellow in the same laboratory working towards his Ph.D. His main interests are analog circuits design, digital image processing
References (24)
- et al.
Automatic signature verification and writer identification – The state of the art
Pattern Recognition
(1989) Recognition of writer by handwriting images
Pattern Recognition
(1981)- et al.
Signature verification using global and grid features
Pattern Recognition
(1994) Grayscale morphology, Comput. Vision Graphics Image Process.
(1986)- et al.
A neural network approach to off-line signature verification using directional pdf
Pattern Recognition
(1996) - et al.
On a formal measurement of individual variability appeared in handwritten characters
Trans. IECE Japan
(1980) - E.N. Zois, V. Anastassopoulos, Fusion of correlated decisions for writer identification, Proceedings of the IEE Third...
- V. Clement, K. Steinke, R. Naske, The application of image proccesing and writer recognition techniques to the forensic...
- V. Clement, Forensic writer recognition, Proceedings of the NATO Advanced Studies Institute, 1981, pp....
- E.N. Zois, V. Anastassopoulos, Methods for writer identification, Proceedings of the International Conference on...
Vision in Man and Machine
Cited by (129)
DCWI: Distribution descriptive curve and Cellular automata based Writer Identification
2019, Expert Systems with ApplicationsArt Image Identification Based on Feature Recognition and BP Neural Network
2024, Advances in Transdisciplinary EngineeringShort Words for Writer Identification Using Neural Networks
2023, Applied Sciences (Switzerland)Short Words Signature Verification using Markov Chain and Fisher Linear Discriminant Approach
2022, International Journal of Advanced Computer Science and ApplicationsA revised range of variability approach considering the morphological alteration of hydrological indicators
2021, Stochastic Environmental Research and Risk Assessment
About the Author—ELIAS ZOIS was born in Athens, Greece in 1971. In 1994 he received his B.Sc. degree in Physics from Physics Department, University of Patras, Greece. In 1997 he received his M.Sc. in Electronics from Electronics Laboratory, University of Patras, Greece. His Master Thesis was on handwritten text analysis and writer identification. He is now a research fellow in the same laboratory working towards his Ph.D. His main interests are analog circuits design, digital image processing and pattern recognition and classification with emphasis on handwritten text analysis and discrimination.
About the Author—VASSILIS ANASTASSOPOULOS was born in Patras, Greece, in 1958. He received the B.Sc. degree in Physics in 1980 from the University of Patras, Greece and the Ph.D. in Electronics in 1986 from the same University. His Ph.D. Thesis was on Digital Signal Processing and in particular, on Delta Modulation Filters.
From 1980 to 1985 he was employed as a research assistant in the Electronics Laboratory, University of Patras. From 1985 to 1989 he was a lecturer in the same Laboratory. From 1989 to 1990 he worked as a research associate in the Department of Electrical Engineering, University of Toronto, on Nonlinear Filters and Pattern Recognition and Classification Techniques. From 1992 he is an Assistant Professor in Electronics Laboratory, University of Patras, Greece. Since 1990, he has been in close cooperation with AUG Signals Ltd, a Canadian company, working on Radar Signal Detection, IR Image Processing and Data Fusion. He worked with this company during his sabbatical in Canada (1994–1995).
His research interests are within the scope of Digital Signal Processing, Image Processing, Radar Signal Processing and Pattern Recognition and Classification. He is a member of the IEEE.