Robust soft-biometrics prediction from off-line handwriting analysis

Currently, writer’s soft-biometrics prediction is gaining an important role in various domains related to forensics and anonymous writing identification. The purpose of this work is to develop a robust prediction of the writer’s gender, age range and handedness. First, three prediction systems using SVM classifier and different features, that are pixel density, pixel distribution and gradient local binary patterns, are proposed. Since each system performs differently to the others, a combination method that aggregates a robust prediction from individual systems, is proposed. This combination uses Fuzzy MIN and MAX rules to combine membership degrees derived from predictor outputs according to their performances, oft-biometrics uzzy MIN-MAX combination LBP VM which are modeled by Fuzzy measures. Experiments are conducted on two Arabic and English public handwriting datasets. The comparison of individual predictors with the state of the art highlights the relevance of proposed features. Besides, the proposed Fuzzy MIN-MAX combination comfortably outperforms individual systems and classical combination rules. Relatively to Sugeno’s Fuzzy Integral, it has similar computational complexity while performing better in most cases. © 2015 Elsevier B.V. All rights reserved. 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 . Introduction Handwriting recognition plays essential roles in various life domains such s mail sorting and bank checks verification. With the new technologies it is ncreasingly sought in more specific applications including information retrieval in istorical documents and soft-biometrics prediction. Soft-biometrics is all what our enses perceive to differentiate us from each others, such as the age range, eye color, ender and ethnicity. These constitute key demographic attributes, which help to lassify the human being into categories. During the last years, soft-biometrics traits ere systematically predicted from face images [1,2]. Currently, there is a significant umber of organizations that already employ handwriting analysis for personalty profiling [3,4]. In fact, either for forensic identification of anonymous writing uthor, or the attribution of historical handwritten documents, soft-biometrics can e extremely useful. Furthermore, various studies tried to explain how the gender an control the human behavior. Specifically, the gender impact has been proved in arkinson disease [5], motor learning [6], dichotic listening [7] as well as in crimes nd violence [8]. Therefore, researchers in handwriting recognition were faced to a traightforward question, that is: Can the gender and other soft-biometrics influence he handwriting? In [9], authors investigated the relationship between sex horones and the handwriting style. Their findings showed that prenatal sex hormones an affect the women handwriting. In some earlier psychological investigations, diferences between men’s and women’s handwriting were examined [10,11]. Besides, n [12], experts were asked to predict the writer’s gender from handwritten senPlease cite this article in press as: N. Bouadjenek, et al., Robust soft-bio Comput. J. (2015), http://dx.doi.org/10.1016/j.asoc.2015.10.021 ences. Experiments reported prediction accuracy about 68%. Also, in [13–15], age mpact over the handwriting performance was investigated while in [16,17], authors ried to highlight the relationship between handedness and language dominance. ∗ Corresponding author. Tel.: +213 551544573. E-mail addresses: [email protected] (N. Bouadjenek), [email protected] (H. Nemmour), [email protected] (Y. Chibani). ttp://dx.doi.org/10.1016/j.asoc.2015.10.021 568-4946/© 2015 Elsevier B.V. All rights reserved. 68 69 70 71 72 In handwritten document analysis field, automatic soft-biometrics prediction constitutes a new research subject. The literature reports only few studies, which addressed gender, handedness, age range and nationality prediction. The first work was developed in 2001 by Cha et al. [18]. Thereafter some other works were reported in [19–22]. A prediction system is composed of two main steps, that are feature generation and classification. In each step, efficient methods are required to achieve satisfactory performance. The key idea for developing a handwriting recognition system, is the choice of feature generation and classification schemes. Regarding the recognition step, a large number of classifiers that are based on different concepts such as singular value decomposition, principle component analysis, statistical modeling, as well as support vector methods are widely used for handwriting recognition [23]. In previous works on soft-biometrics prediction, various robust classifiers such as neural networks, SVM and Random Forests were employed while the feature generation was based on conventional direction, curvature and edge features. Note that SVM are considered as the best choice in most of recognition tasks where they commonly outperform other learning machines, namely, neural networks and HMM [24,25]. In fact, SVM are based on structural risk minimization, which answers two main problems of the statistical learning theory that are overfitting and controlling the classification complexity [26]. In addition, their training formulation is perfectly adequate to handle data with very large size without requiring dimensionality reduction. Furthermore, gender prediction results reported in [27] reveal that one SVM classifier could outperform the combination of multiple systems if they employ weak descriptors. Therefore, a straightforward way to achieve an efficient prediction is to associate robust data features to SVM. This work is focused on the use of effective topological and gradient features, which are more suitable for handwriting characterization. Considered topological features are pixel density and pixel distribution, which gave satisfactory permetrics prediction from off-line handwriting analysis, Appl. Soft formance in handwritten signature verification [28]. As gradient feature the so called Gradient Local Binary Patterns (GLBP) is used. This descriptor was recently introduced for human detection in order to improve the histogram of oriented gradients. Three SVM predictors based respectively, on pixel density, pixel distribution and GLBP, are developed. Subsequently, a Fuzzy MIN-MAX combination algorithm 73 74 75 76 77