 Research
 Open Access
A novel approach for handedness detection from offline handwriting using fuzzy conceptual reduction
 Somaya AlMaadeed^{1}Email author,
 Fethi Ferjani^{1},
 Samir Elloumi^{1} and
 Ali Jaoua^{1}
https://doi.org/10.1186/s136400150097y
© AlMaadeed et al. 2015
 Received: 25 September 2014
 Accepted: 27 November 2015
 Published: 5 January 2016
Abstract
A challenging area of pattern recognition is the recognition of handwritten texts in different languages and the reduction of a volume of data to the greatest extent while preserving associations (or dependencies) between objects of the original data. Until now, only a few studies have been carried out in the area of dimensionality reduction for handedness detection from offline handwriting textual data. Nevertheless, further investigating new techniques to reduce the large amount of processed data in this field is worthwhile. In this paper, we demonstrate that it is important to select only the most characterizing features from handwritings and reject all those that do not contribute effectively to the process of handwriting recognition. To achieve this goal, the proposed approach is based mainly on fuzzy conceptual reduction by applying the Lukasiewicz implication. Handwritten texts in both Arabic and English languages are considered in this study. To evaluate the effectiveness of our proposal approach, classification is carried out using a KNearestNeighbors (KNN) classifier using a database of 121 writers. We consider left/right handedness as parameters for the evaluation where we determine the recall/precision and Fmeasure of each writer. Then, we apply dimensionality reduction based on fuzzy conceptual reduction by using the Lukasiewicz implication. Our novel feature reduction method achieves a maximum reduction rate of 83.43 %, thus making the testing phase much faster. The proposed fuzzy conceptual reduction algorithm is able to reduce the feature vector dimension by 31.3 % compared to the original “best of all combined features” algorithm.
Keywords
 Index termshandwriting
 Fuzzy binary relation
 Left/right identification
 Feature
 Lukasiewicz implication
 Galois connections
 Closure of Fuzzy Galois connections
1 Introduction
Handwriting recognition is the ability of a computer to receive and interpret intelligible handwritten input from sources such as paper documents, photographs, touchscreens, and other devices. The image of the written text may be sensed “offline” from a piece of paper by optical scanning (optical character recognition) or intelligent word recognition. Alternatively, the movements of the pen tip may be sensed “online”, for example by a penbased computer screen surface. Handwriting recognition principally entails optical character recognition. However, a complete handwriting recognition system also handles formatting, performs correct segmentation into characters and determines the most plausible words.
Handwriting recognition has been one of the fascinating and challenging research areas in the field of image processing and pattern recognition [1, 2]. It contributes immensely to the advancement of an automation process and can improve the interface between human beings and machines in numerous applications. Several research works have been focusing on new techniques and methods that would reduce the processing time while providing higher recognition accuracy [3].
In general, handwriting recognition is classified into two types of methods: offline and online handwriting recognition methods. In offline recognition, the writing is usually captured optically by a scanner, and the complete text is available as an image. In the online system, the twodimensional coordinates of successive points are represented as a function of time, and the order of strokes made by the writer is also available. The online methods have been shown to be superior to their offline counterparts in recognizing handwritten characters due to the temporal information available with the former [4, 5]. Several applications, including mail sorting, bank processing, document reading, and postal address recognition, require offline handwriting recognition systems. As a result, offline handwriting recognition continues to be an active area of research toward exploring newer techniques that would improve recognition accuracy [6].
In our current study, we focus on offline handwriting and use a KNearest Neighbors (KNN) classifier to make classifications based on many parameters, such as gender, age, handedness, and nationality, to measure the performance of our proposed algorithm. This type of classification has several applications. For example, in the forensic domain, handwriting classification can help investigators to focus on a certain category of suspects. Additionally, processing each category separately leads to improved results in writer identification and verification applications.
There are few studies in the literature that investigate the automatic detection of gender, age, and handedness from handwritings. Bandi et al. [7] proposed a system that classifies handwritings into demographic categories using the “macrofeatures” introduced in [8]. These features focus on measurements such as pen pressure, writing movement, stroke formation, and word proportion. The authors reported classification accuracies of 77.5, 86.6, and 74.4 % for gender, age and handedness classification, respectively. However, in this study, all the writers produced the same letter. Unfortunately, this is not always the case in real forensic caseworks. Moreover, the dataset used in this study is not publicly available.
Liwicki et al. [9] also addressed the classification of gender and handedness in the online mode (which means that the temporal information about the handwriting is available). The authors used a set of 29 features extracted from both online information and its offline representation and applied support vector machines and Gaussian mixture models to perform the classification. The authors reported a performance of 67.06 % for gender classification and 84.66 % for handedness classification. In [10], the authors separately reported the performance of the offline mode, the online mode, and their combination. The performance reported for the offline mode was 55.39 %, which is slightly better than chance.
In this paper, we propose a novel approach to the detection of the handedness of the writer of a handwritten document based on the Lukasiewicz implication where a set of features was proposed and evaluated to predict the handedness of the writer. These features are combined using a KNearest Neighbors (KNN) classifier under the Rapidminer platform [11]. This method is evaluated using the QUWI database, which is the only publicly available dataset containing annotations regarding gender, age range, and nationality.
The rest of the paper is organized as follows. In section 2, we review some basic definitions from relational algebra, the mathematical background related to fuzzy set theories and useful for this research paper. In section 3, the stateoftheart in writer identification for the English and Arabic languages is presented in detail. The evaluation was made using larger amounts of text and may not produce acceptable results when limited amounts of text are available. Writer recognition from short handwritten texts is therefore an interesting area of study. Section 4 gives a description of the system overview and the database used for carrying out the experimental evaluations. Next, we describe the proposed features and the method in which they are extracted. Then, we present the utilized (KNN) classifier followed by the detailed results and an analysis of the experimental evaluations. Finally, we conclude the paper with some discussion on future research directions on the subject.
2 Key settings and new definitions
The domains of computer science, relational algebra, formal concept analysis, and lattice theory have seen important advances in research [12]. This research has contributed enormously to the search for original solutions for complex problems in the domains of knowledge engineering, data mining, and information retrieval. Relational algebra and formal concept analysis may be considered as useful mathematical foundations that unify data and knowledge in information retrieval systems.
2.1 Binary relations

A relation \(\mathcal {R}\) is a subset of the Cartesian product of two sets \(\mathcal {X}\) and \(\mathcal {Y}\).

An element \((e,e') \in \mathcal {R}\), where e ^{′} denotes the image of e by \(\mathcal {R}\).

A binary relation identity \(\mathcal {I}(\mathcal {A})= \{ (e,e)e \in \mathcal {A} \}\).

The relative product or composition of two binary relations \(\mathcal {R}\) and \(\mathcal {R}^{'}\) is \(\mathcal {R} \circ \mathcal {R}^{'} = \{(e,e')  \exists t \in \mathcal {Y} : ((e,t) \in \mathcal {R}) \& ((t,e') \in \mathcal {R}') \}\).

The inverse of the relation \(\mathcal {R}\) is \(\mathcal {R}^{1} = \{(e, e')  (e', e) \in \mathcal {R}\}\).

The set of images of e is defined by \(e.\mathcal {R} = \{e'  (e, e') \in \mathcal {R}\}\).

The set of antecedents of e ^{′} is defined by \(\mathcal {R}.e' = \{e  (e, e') \in \mathcal {R}\}\).

The cardinality of \(\mathcal {R}\) is defined by \(Card(\mathcal {R})\) = the numbers of pairs \((e,e') \in \mathcal {R}\).

The complement of the relation \(\mathcal {R}\) is \(\overline {\mathcal {R}}= \{ (e,e')  (e,e') \notin \mathcal {R}\}\).

The domain of \(\mathcal {R}\) is defined by \(Dom(\mathcal {R}) = \{e \exists e':(e,e') \in \mathcal {R}\}\).

The range or codomain of \(\mathcal {R}\) is defined by \(Cod(\mathcal {R}) = \{e'  \exists e :(e,e') \in \mathcal {R}\}\).
2.2 Formal concept analysis
Formal Concept Analysis (FCA) is the mathematical theory of data analysis using formal contexts and concept lattices [12–14]. It was introduced by Rudolf Wille in 1984 and builds on applied lattice and order theory, which were developed by Birkhoff et al. [15]
Definition 1.
A formal context
A formal context (or an extraction con text) is a triplet \(\mathcal {K} = (\mathcal {X},\mathcal {Y},\mathcal {R})\), where \(\mathcal {X}\) represents a finite set of objects, \(\mathcal {Y}\) is a finite set of attributes (or properties), and \(\mathcal {R}\) is a binary (incidence) relation, (i.e., \(\mathcal {R} \subseteq \mathcal {X}\times \mathcal {Y})\). Each couple \((x,y)\in \mathcal {R}\) expresses that the object \(x\in \mathcal {X}\) verifies property y belonging to \(\mathcal {Y}\).
Definition 2.
Formal concept in fuzzy binary relation
Let \(\mathcal {X}\) be a set called the universe of discourse. Elements of \(\mathcal {X}\) are denoted by lowercase letters. A fuzzy set E ={ x _{1}/v _{1},x _{2}/v _{2},⋯,x _{ n }/v _{ n }} is defined as a collection of elements \(x_{i} \in \mathcal {X}, i=1:n\), which includes a degree of membership v _{ i } for each element x _{ i }[16,17].
A fuzzy binary context (or fuzzy binary relation) is a fuzzy set defined on the product of two sets \(\mathcal {O}\) (set of objects) and \(\mathcal {P}\) (set of properties). Hence, \(\mathcal {X}=\mathcal {O} \times \mathcal {P}\).
Definition 3.
Galois connection
 1.
\(f(A) = \left \{e'  \forall e \in A, (e,e') \in \mathcal {R}\right \},\)
 2.
\(g(B) = \left \{e  \forall e' \in B, (e,e') \in \mathcal {R}\right \}.\)
Operator f defines the properties shared by all elements of A, and operator g defines objects sharing the same properties included in set B. The operators f and g define a Galois connection between the sets \(\mathcal {X}\) and \(\mathcal {Y}\) with respect to the binary context \((\mathcal {X},\mathcal {Y},\mathcal {R})\)[12,18].
Definition 4.
Fuzzy set

μ _{ A } : \(\mathcal {X} \rightarrow [0, 1]\),

A finite fuzzy set can be denoted as
A = {μ _{ A }(x _{1})/x _{1},μ _{ A }(x _{2})/x _{2},…,μ _{ A }(x _{ n })/x _{ n }},
for any \(x_{i} \in \mathcal {X}\).
Example 1.
Fuzzy binary relation
a  b  c  d  e  f  

O _{1}  0.5  0.2  0.6  0.4  0.7  0.5 
O _{2}  0.7  0.3  0.2  0.3  0.2  1 
O _{3}  1  0.4  0.6  1  0.7  0.5 
O _{4}  0.5  0.2  0.6  0.4  0.8  0.6 
O _{5}  0.7  0.3  0.7  0.4  0.6  0.7 
Definition 5.
Fuzzy Galois connection

\(\mathcal {F}(A) = \left \{d / \alpha  \alpha = min\{\mu _{\mathcal {F}_{\textit {BR}}} (g,d)  g \in A, d \in \mathcal {P}\}\right \}\)

\(\mathcal {H}_{\delta }(B) = \left \{g  d \in \mathcal {P} \Rightarrow (\mu _{\mathcal {B}}(d) \rightarrow _{L} \mu _{\mathcal {F}_{\textit {BR}}} (g,d) \geqslant \delta \right \},\)
Note that \(\mu _{\mathcal {F}_{\textit {BR}}}(g,d)\)denotes the weight of the pair (g,d) in the fuzzy relation \(\mathcal {F}_{\textit {BR}}\).
Definition 6.
A fuzzy closure operator
For two sets A and B such that \(A \subseteq \mathcal {O}\), B is a fuzzy set defined on \(\mathcal {P}\), and δ∈ [ 0,1]. We define \(Closure(A) = \mathcal {H}_{\delta }(\mathcal {F}(A))= A'\) and \(Closure(B) = \mathcal {F}(\mathcal {H}_{\delta }(B))= B'\).
 1.
\(A_{i} \subseteq A_{j} \Rightarrow f(A_{i}) \supseteq f(A_{j})\);
 2.
\(B_{i} \subseteq B_{j} \Rightarrow g(B_{i}) \supseteq g(B_{j})\);
 3.
\(A_{i} \subseteq g \circ f(A_{i})\) and \(B_{i} \subseteq f \circ g(B_{i})\);
 4.
\(A \subseteq g(B) \Leftrightarrow B = f(A)\); and
 5.
f=f∘g∘f and g=g∘f∘g.
Fuzzy data reduction To manage the large amount of features, it is important to select the most pertinent ones. In this paper, we use fuzzy conceptual reduction applied to the original data. Fuzzy conceptual data reduction methods have the main objective of minimizing the size of data while preserving the content of the original document. Unfortunately, most of the methods presented in the literature are based on heuristics and are not accurate. Moreover, reducing fuzzy data becomes a difficult problem because the handling of imprecision and uncertainty may cause information loss and/or deformation. In this work, we develop a fuzzy conceptual approach based on Lukasiewicz fuzzy Galois. This method is based on fuzzy formal concept analysis, which has been recently developed by several researches and applied for learning, knowledge acquisition, information retrieval, etc. The Lukasiewicz implication based on the fuzzy Galois connection is mainly used in this paper. It allows one to consider different precision levels according to the value of δ in the definition of fuzzy formal concepts. The advantage of reduced data is that it can be used directly as a prototype for making decisions, for supervised learning, or for reasoning. For that purpose, we first prove that some rows can be removed from the initial fuzzy binary context at a given precision level (value given to δ by application of the Lukasiewicz implication). It is primordial to assess that there is an equivalence between an object and a set of objects. Second, we define a solution for data reduction in the case of fuzzy binary relations.
Equivalence between an object and a subset of other objects An object x is equivalent (for a given value of δ for δ varying from 0 to 1) to a set of objects S _{ x }, relative to a fuzzy binary context \(\mathcal {F}_{\textit {BR}}\), if and only if \(\{x\} \cup S_{x}\) is a domain of a concept of \(\mathcal {F}_{\textit {BR}}\), and the closure \((S_{x}) = \{x\} \cup S_{x}\), where x∉S _{ x }. As intuitive justification, x is equivalent to S _{ x } means that \(S_{x} \rightarrow x\) within some precision δ.
3 A review of related works
4 System overview of proposed generic approach for combined features
In this section, we present a system overview of our proposed approach. We then describe the dataset that was utilized to obtain the results. In the following, we give a description of the feature extraction and subsequently detail our proposed algorithm. The main experiment is discussed in this paper.
4.1 System overview
4.2 Description of the dataset
4.3 Feature extraction
To make the system independent of the pen, images are first binarized using the Otsu thresholding algorithm [34]. The following subsections describe the features considered in this study. It is to be noted that these features do not correspond to a single value but are defined by a probability distribution function (PDF) extracted from the handwriting images to characterize the writer’s individuality [35,36]. The PDF describes the relative likelihood for a certain feature to take on a given value.
4.3.1 Directions (f1)
We move along the pixels of the obtained segments of the skeleton using a predefined order favoring the four connectivity neighbors. For each pixel p, we consider the 2×N+1 neighboring pixels centered at p. The linear regression of these pixels is calculated to give the tangent at the pixel p [36].
The PDF of the resulting directions is computed as a probability vector for which the size has been empirically set to 10.
4.3.2 Curvatures (f2)
For each pixel p belonging to the contour, we consider a neighboring window, which has a size t. We compute the number of pixels n _{1} inside this neighboring window belonging to the background and the number of pixels n _{2} representing the foreground. Obviously, the difference n _{1}−n _{2} increases with the local curvature of the contour. We then estimate the curvature as being \(C=\frac {n_{1}n_{2}}{n_{1}+n_{2}}\). The PDF of the curvatures is computed as a vector whose size has been empirically set to 100 (s pixels in each side) [36].
4.3.3 Tortuosity (f3)
This feature makes it possible to distinguish between fast writers who produce smooth handwriting and slow writers who produce “ortuous”/twisted handwriting [36] by finding the longest line in the middle of the character shape. This feature has a PDF vector of 10. The PDF of the angles of the longest traversing segments are produced in a vector whose size has been set to 10.
4.3.4 Chain code features (f4f7)

f4: PDF of i patterns in the chain code list such that i∈0,1,…,7. This PDF has a size of 8.

f5: PDF of (i,j) patterns in the chain code list such that i,j∈0,1,…,7. This PDF has a size of 64.

Similarly, f6 and f7 correspond to a PDF of (i,j,k) and (i,j,k,l) in the chain code list, where i,j,k,l∈0,1,…,7. Their respective sizes are 512 and 4096.
4.3.5 Edgebased directional features (f8f26)
Overview of the implemented features
Feature  Description  Dimension 

f1  Runlength distribution of white pixels  10 
in four directions  
f2  Runlength distribution  100 
of black pixels in four directions  
f3  Runlength distribution of white  10 
and black pixels in four directions  
f4  Edgedirection distribution using 16  8 
angles  
f6  Polygonbased features  512 
f16  Chaincodebased local features  36 
f23  Codebookbased features  112 
f26  ARcoefficientbased features  220 
Each contour C o n t o u r _{ i }, being a sequence of consecutive boundary points, is computed as follows:
\({Contour}_{i} = \{p_{j}j \leqslant M_{i}, p_{1}=p_{\textit {Mi}} \}\), where M _{ i } is the length of c o n t o u r _{ i }.
In the following section, we will present the classifier used to predict the class of the set of features. Then, we define the different steps of the proposed algorithm.
4.4 Proposed classifier
4.4.1 (KNN) Classifier
The KNearestNeighbor (KNN) classifier is one of the most basic classifiers for pattern recognition and data classification. The principle of this method is based on the intuitive concept that data instances of the same class should be closer in the feature space. As a result, for a given data point x of an unknown class, we can simply compute the distance between x and all the data points in the training data and assign the class determined by the K nearest points of x. Due to its simplicity, KNN is often used as a baseline method in comparison with other sophisticated approaches utilized in pattern recognition. The KNearestNeighbor classification divides data into a test set and a training set. In our case, we choose K = 5 to be used in a sample of 128 writers for both English and Arabic texts.
The main task of classification is to use the feature vectors provided by the feature extraction algorithm to assign the object to a category [37]. In our work, we use the KNearest Neighbors (KNN) for the classification of the extracted features. KNN running on the Rapidminer platform [11] classifier classifies an unknown sample based on the known classification of its neighbors [38,39]. Given unknown data, the KNearestNeighbor classifier searches the pattern space for the K training data that are closest to the unknown data. These K training tuples are the K “Nearest Neighbors” of the unknown data. Typically, we normalize the values of each attribute. This helps to prevent attributes with initially large ranges from outweighting attributes with initially smaller ranges (such as binary attributes).
In this step, the features previously presented are used to predict the handedness of the writer of each document. When performing the classification, each element of the feature vector will be used as a separate input for the classifier (for example, f1 will be an input vector of 10 elements for the classifier, as shown in Table 2). We have combined these features using a KNearestNeighbor classifier [40]. A description of the combination of features using the (KNN) classifier is given below.
4.4.2 Proposed algorithm
In this section, we give a description of the proposed algorithm based on a new heuristic approach for obtaining the best feature combination by applying the Lukasiewicz implication with variations of different values of δ. Only the best value of δ that provides the highest Fmeasure score was retained. The proposed approach is split into three submodules: (1) the main algorithm, which takes into consideration the input fuzzy binary relation for the features (f1, f2, f3, f4, f6, f16, f23 and f26) denoted \(\mathcal {F}_{\textit {BR}}\) (these features are presented in excel files in the form of fuzzy binary relations, where the rows represent the different writers based on their left/right handedness and the columns represent the values of features); (2) the remaining feature module processing that identifies the features to be rejected and the features that will be maintained according to the best value of δ that provides a high score of the Fmeasure and a considerable improvement in the data reduction percentage (which is accomplished by computing the closure of the Galois connection); and (3) the third submodule, which determines the computed closure of the remaining attributes. In this case, we consider the fuzzy binary matrix relevant to a given feature F _{ i }, where \(\mathcal {O}_{i}\) represents an object and A _{0},A _{1},…,A _{ n } represent the corresponding attributes. The rows represent the different writers (left and right handedness). We use 121 writers. The columns represent the measured values of the features. For instance, feature f26 describes the measured values of the "Polygonbased features" using 512 measures (i.e., 0.019815 represents a measure).
Step 1: Performing a matrix transposition; transpose the rows to the attributes and columns to the objects. Each row represents a different feature F _{ i }, and the objects corresponding to 121 different writers with both left and right handedness are represented as columns.
Step 2: Choosing different arbitrary values of δ, we compute the closure of the attributes by using the following formula: h _{ δ }∘f(A)={A _{0},A _{1},…,A _{ n }}. The discovered redundant attributes are removed. Intuitively, an attribute is redundant if we can regenerate it by association from other attributes. Finally, we keep the last subset that contains the reduced subset of O b j e c t s×A t t r i b u t e s with the highest values of δ in terms of precision, recall and Fmeasure.
 1.In the main algorithm, we determine for each row

The closure list “ListClose”, which is denoted S _{ x } and computed using the following formula: \(\mathcal {H} \circ \mathcal {F}(x)\).

The next step consists of removing from Listclose the redundant feature: The values of \(\mathcal {F}_{\mathcal {BR}} (i.e., \textsc {CL}_{\textit {Sx}} \leftarrow \textsc {ListClose.closure}  (x\)).

If C L _{ Sx } is equivalent to S _{ x }, then feature x is removed.

 2.
The second submodule (Algorithm 2) consists of computing the Galois connection of L i s t S _{ x } according to the specified value of δ (i.e., CL_{ Sx })
 3.
Another module may be added in order to update the context if C L S _{ x } is equal to S _{ x }.

Compute the Fmeasure for each feature.
 1.
Vary different values of δ (for example, 0.95), and generate the features
that satisfy the Lukasiewicz implication according to the fixed value of δ;
 2.
Choose the best results of the features and determine the percentage of reduction;
 3.
Combine the feature with the highest score (e.g., the Fmeasure) with all the other features;
 4.
Compute the Fmeasure for the combined two selected features; and
 5.
Retain only the combined feature with the highest score.
 1.

Repeat the steps above in a similar way for combinations of the next levels (3, 4 and so on) until no improvement is obtained.

Select the combination of features with the highest Fmeasure score.
4.5 Results and their analysis

α _{1}=# of true documents correctly classified into C l a s s _{ i }

α _{2}=# of true documents incorrectly classified into C l a s s _{ i }

β _{1}=# of false documents correctly classified into C l a s s _{ i }

β _{2}=# of false documents incorrectly classified into C l a s s _{ i }
The precision, recall and Fmeasure metrics are used to evaluate our approach using English and Arabic texts. In this work, we have chosen the KNearestNeighbors (KNN) algorithm, which is widely used for classification, machine learning, and pattern recognition by data miners [42].

Divide training examples into two sets, a training set (95 %) and a validation set (5 %);

Predict the class labels for the validation set by using the examples in the training set; and

Choose the number of neighbors K = 5 that maximizes the classification accuracy.
4.5.1 Experiments with no feature reduction
The classification is carried out separately for the Arabic and English languages in a first step and jointly in a second step. The results are reported for the case of similar texts written by all the writers and different texts for each writer. In the following, we present the results of the classification at the end of each iteration:
First iteration: we compute the Fmeasure of the features separately. We present (1) the reduction percentage and (2) the improvement of the Fmeasure through application of the Lukasiewicz implication.
Detailed accuracy for left and right handwriting: fuzzy 2combinations
Left and right handedness  

# of Att  δ  % Reduction  Precision  Recall  Fmeasure 
f1f2  0.995  17.7 %  65.69 %  66.86 %  66.27 % 
f1f3  0.85  31.3 %  69.47 %  69.40 %  69.43 % 
f1f4  0.92  20.0 %  62.81 %  62.80 %  62.80 % 
f1f23  0.995  34.3 %  64.55 %  64.43 %  64.48 % 
f1f6  0.995  5.9 %  63.55 %  61.93 %  62.73 % 
f1f26  0.998  36.9 %  67.23 %  66.89 %  67.05 % 
f1f16  0.993  23.8 %  63.85 %  63.58 %  63.71 % 
Detailed accuracy for left and right handwriting: fuzzy 3combinations
Left and right handedness  

# of Att  δ  % Reduction  Precision  Recall  Fmeasure 
f1f3f2  0.98  32.3 %  67.81 %  67.75 %  67.77 % 
f1f3f4  0.94  16.7 %  70.27 %  70.26 %  70.26 % 
f1f3f23  0.995  33.3 %  71.99 %  71.87 %  71.93 % 
f1f3f6  0.997  9.9 %  68.91 %  68.58 %  68.59 % 
f1f3f26  0.997  19.6 %  71.99 %  71.87 %  71.93 % 
f1f3f16  0.994  24.4 %  71.99 %  71.87 %  71.93 % 
Detailed accuracy for left and right handwriting: fuzzy 5combinations
Left and right handedness  

# of Att  δ  % Reduction  Precision  Recall  Fmeasure 
f1f3f23f16f2  0.995  8.9 %  66.94 %  66.94 %  66.94 % 
f1f3f23f16f4  0.992  32.4 %  66.12 %  66.12 %  66.12 % 
f1f3f23f16f6  0.990  11.7 %  70.27 %  70.23 %  70.25 % 
f1f3f23f16f26  0.995  31.3 %  71.99 %  71.87 %  71.93 % 
Detailed accuracy for left and right handwriting: fuzzy 6combinations
Left and right handedness  

# of Att  δ  % Reduction  Precision  Recall  Fmeasure 
f1f3f23f16f26f2  0.990  29.7 %  66.15 %  66.09 %  66.12 % 
f1f3f23f16f26f4  0.997  18.8 %  66.97 %  66.95 %  66.96 % 
f1f3f23f16f26f6  0.992  20.6 %  69.47 %  69.40 %  69.43 % 
Detailed accuracy for left and right handwriting: fuzzy 7combinations
Left and right handedness  

# of Att  δ  % Reduction  Precision  Recall  Fmeasure 
f1f3f23f16f26f4f2  0.995  4.6 %  67.01 %  66.91 %  66.96 % 
f1f3f23f16f26f4f6  0.990  30.7 %  66.94 % 
Detailed accuracy for left and right handwriting: fuzzy 8combinations
Left and right handedness  

# of Att  δ  % Reduction  Precision  Recall  Fmeasure 
f1f3f23f16f26f4f6f2  0.995  5.2 %  68.77 %  68.55 %  68.66 % 
Detailed accuracy for left and right handwriting: fuzzy 4combinations
Left and right handedness  

# of Att  δ  % Reduction  Precision  Recall  Fmeasure 
f1f3f23f2  0.995  12.1 %  66.96 %  66.93 %  66.94 % 
f1f3f23f4  0.992  46.8 %  67.02 %  66.97 %  66.99 % 
f1f3f23f6  0.995  33.3 %  71.99 %  71.87 %  71.93 % 
f1f3f23f26  0.997  16.2 %  71.99 %  71.87 %  71.93 % 
f1f3f23f16  0.992  35.0 %  71.99 %  71.87 %  71.93 % 
The combined features f1f3f23f16f26 yield the highest score (Fmeasure is approximately equivalent to 71.93 %). Therefore, this feature will be combined with the other features (i.e., f26f4, f26f23, f26f16, f26f3, f26f6, and f261). We continue the combination process until reaching the point where no possible combination of features that obtain a higher Fmeasure score is possible. In the previous tables, this is shown by the recognition rates and the reduction rates which start to decrease. Finally, the Fmeasure slightly improves with the 8combination with only a 5 % reduction rate.
4.5.2 Summary of experiments using the Lukasiewicz implication
4.5.3 Key findings
In the following, we provide a summary of the obtained results. We show the best results for each combination. We then graph these results in an appropriate figure. Finally, we comment on the results.

The accuracy is reasonable (69.78 % on average) in all approaches except for the combination f1f3 (69.47 %). This is due to the quality of the features (the accuracy of f1 was 70.58 %, and the accuracy of f3 was 62.85 %). It reaches its maximum for the following combination (approximately 71.99 %): f1f3f23f16f26 (Table 10).Table 10
Different combinations
Left and right handedness
Combination
Features selected
1combination
f6
2combination
f1f3
3combination
f1f3f23
4combination
f1f3f23f16
5combination
f1f3f23f16f26
6combination
f1f3f23f16f26f6
7combination
f1f3f23f16f26f4f2
8combination
f1f3f23f16f26f4f6f2

The recall attains its highest values (71.87 %) for the combination f1f3f23f16 and the lowest value (69.60 %) for the combination f1f3f23f16f26f4f2. It is clear that the features f26f4f2 did not improve the score. On average, the score was 69.60 %.

The Fmeasure reaches its highest values (71.93 %) for the combination f1f3f23f16 and the lowest value (69.69 %) for the combination f1f3f23f16f26f4f2. It is clear that the same features f26f4f2 did not improve the score. On average, the score was 69.69 %.

The reduction percentage reaches its maximum (83.43 %) for the feature F6 alone and its minimum of 4.57 % for the feature f1f3f23f16f26f4f2, while on average, the reduction percentage was 30.36 %); and

Finally, if one takes into consideration the highest Fmeasure with an improvement in the reduction percentage, it is clear that using f1f3f23f16 with the Fmeasure (71.93 %) results in a percent reduction of 31.3 %, which emphasized a considerable improvement in dimensionality reduction of the features. Tables 11, 12 and 13.Table 11
Detailed accuracy for left and righthandwriting: fuzzy 1combination
Left and right handedness
# of Att
δ
% Reduction
Precision
Recall
Fmeasure
f26
0.997
74.6 %
66.12 %
66.12 %
66.12 %
f6
0.992
83.43 %
70.07 %
69.35 %
69.71 %
f4
0.92
33.3 %
68.67 %
68.57 %
68.61 %
f3
0.851
33.3 %
62.85 %
62.83 %
62.84 %
f16
0.998
5.6 %
59.55 %
59.47 %
59.50 %
f1
0.885
20.0 %
70.58 %
70.19 %
70.73 %
f23
0.999
13.4 %
68.00 %
67.72 %
67.86 %
f2
0.995
32.6 %
45.40 %
45.43 %
45.41 %
Table 12Detailed accuracy for left and right handwriting using the Lukasiewicz implication
Left and right handedness
# of Att
δ
% Reduction
Precision
Recall
Fmeasure
f1f3
f23f16
0.995
31.3 %
71.99 %
71.87 %
71.93 %
f26
Table 13Summary of the highest obtained results using different combinations
Left and right handedness
Combination
δ
% Reduction
Precision
Recall
Fmeasure
1combination
0.885 %
20 %
70.58 %
70.19 %
70.73 %
2combination
0.85 %
31.3 %
69.47 %
69.40 %
69.43 %
3combination
0.995 %
33.3 %
71.99 %
71.87 %
71.93 %
4combination
0.992 %
35 %
71.99 %
71.87 %
71.93 %
5combination
0.995 %
31.3 %
71.99 %
71.87 %
71.93 %
6combination
0.992 %
20 %
66.47 %
66.40 %
69.43 %
7combination
0.995 %
4.6 %
67.01 %
66.91 %
66.96 %
8combination
0.995 %
5.2 %
68.77 %
68.55 %
68.66 %
4.6 Computational complexity reduction
We consider the features f1, f2, f3, f4, f6, f16, f23, and f26 for the computation of the complexity analysis of our approach. Thus, the computational complexity of the determination of the best combined features, for the previous n features, is determined as follows, as the objects correspond to features and attributes to writers. So, the time complexity is definitively \(\mathcal {O}(n \star m^{2})\) where n is the number of writers and m the number of features. As a matter of fact, we are calculating the closure of each one of the m features, where the closure requires m x n comparisons, where n is the number of writers.
5 Conclusion and future work
We have proposed a new generic approach for combined feature extraction based on a successive combination of the best feature with the highest score with every other feature. We plan to apply this approach to many applications including gender, age, and nationality prediction. The goal of this research consists of investigating the textindependent identification of a script writer. We have employed a set of features (e.g., f1, f2, f3, f4, f6, f16, f23, and f26), which have shown promising results on a database of handwritten documents in two different languages: Arabic and English. The evaluations were carried out on the only existing database of its type, containing short writing samples from 121 different writers. The results of the determination of the best combined feature identification are very encouraging (the Fmeasure is approximately equal to 71.08 %). They reflect the effectiveness of the runlength features in a textindependent script environment and validate the hypothesis put forward in this research, i.e., that the writing style remains approximately the same across different scripts. It is also worth mentioning that, unlike most of the studies that use complete pages of text, our results are based on a limited amount of handwritten text, which is more realistic. Another interesting aspect of this study was the evaluation and comparison of a number of stateoftheart methods on this dataset. The features used in these methods naturally show a decrease in performance when exposed to different script scenarios. In all cases, the runlength features outperform these features. Finally, for the comparison of the proposed method with other methods, the average correct handedness detection results are over 83.43 %, which exceeds the results reported in [30] for offline gender identification (70 %) on the same dataset. The results also compare well with the 73 %; 55.39 % reported for gender classification in [36,43] on different datasets. It would be interesting to evaluate these features on a larger dataset with a large number of writers and many scripts per writer. This, however, involves the challenging task of finding individuals who are familiar with multiple scripts. To extend this study, we intend to utilize a database including writing samples in Arabic, French, and other languages provided by the same writer. In addition, classifiers other than those discussed in this paper can be evaluated to find out how they perform in a manyscript environment. The proposed approach can also be extended to include a rejection threshold to reject writers that are not a part of the database. Finally, it would be interesting to apply a feature selection strategy to reduce the dimension of the proposed feature set and to study which subset of features is the most discriminative in characterizing the writers. The application of Latent Semantic Analysis (LSA) techniques seems promising regarding the reduction of the volume of data. These aspects will constitute the focus of our future research on writer recognition. With regard to data reduction, in our future work, we would like to investigate reducing the highdimensional data of features gathered from usercognitive loads, which results from the density of data to be visualized and mined, and reducing the dimensionality of the dataset while associations (or dependencies) between objects as applied to writer identification. This dimensionality reduction will be based on fuzzy conceptual reduction through the application of the Lukasiewicz implication.
6 Endnote
Declarations
Acknowledgments
This publication was made possible by a grant from the Qatar National Research Fund NPRP 098641128. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the QNRF.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Authors’ Affiliations
References
 A Hassaine, S Al Maadeed, J Aljaam, A Jaoua. ICDAR 2013 Competition on Gender Prediction from Handwriting, Twelfth International Conference on Document Analysis and Recognition ICDAR2013 (IEEEWashington, DC,2013). https://www.computer.org/csdl/abs.html.
 S Impedovo, L Ottaviano, S Occhinegro, Optical character recognition. Int.J. Pattern Recognit. Artif. Intell. 5(12), 1–24 (1991).View ArticleGoogle Scholar
 VK Govindan, AP Shivaprasad, Character recognition—a review. Pattern Recognit.23(7), 671–683 (1990).View ArticleGoogle Scholar
 R Plamondon, SN Srihari, Online and offline handwritten character recognition: a comprehensive survey. IEEE Trans. Pattern Anal. Mach. Intell. 22(1), 63–84 (2000).View ArticleGoogle Scholar
 N Arica, F YarmanVural, An overview of character recognition focused on offline handwriting. IEEE Trans. Syst. Man Cybernet. Part C: Appl. Rev. 31(2), 216–233 (2001).View ArticleGoogle Scholar
 U Bhattacharya, BB Chaudhuri, Handwritten numeral databases of Indian scripts and multistage recognition of mixed numerals. IEEE Trans. Pattern Anal. Mach. Intell. 31(3), 444–457 (2009).View ArticleGoogle Scholar
 K Bandi, SN Srihari, in Proceedings of the International Graphonomics Society Conference (IGS). Writer demographic identification using bagging and boosting (Publisher International Graphonomics Society (IGS), 2005), pp. 133–137. http://www.graphonomics.org/publications.php.
 S Srihari, SH Cha, H Arora, S Lee, in Proceedings of the Sixth International Conference on Document Analysis and Recognition. Individuality of handwriting: a validation study (IEEE, 2001), pp. 106–109.Google Scholar
 M Liwicki, A Schlapbach, P Loretan, H Bunke, in Proceedings of the 13th Conference of the International Graphonomics Society. Automatic detection of gender and handedness from online handwriting (Publisher International Graphonomics Society (IGS), 2007), pp. 179–183. http://www.graphonomics.org/publications.php.
 M Liwicki, A Schlapbach, H Bunke, Automatic gender detection using online and offline information. Pattern. Anal. Appl. 14:, 87–92 (2011).View ArticleMathSciNetGoogle Scholar
 KDnuggets, Data integration, analytical ETL, data analysis, and reporting, rapid miner journal (2012). Software available at, http://sourceforge.net/projects/rapidminer/.
 B Ganter, R Wille, Formal Concept Analysis (SpringerVerlag, Berlin Heidelberg, 1999).View ArticleMATHGoogle Scholar
 L Wang, in Proceedings Part I of the Second International Conference, (FSKD 2005), Changsha, China. Fuzzy Systems and knowledge discovery (SpringlerVerlag BerlinHeidelberg, 2005), pp. 515–519. ISBN 10 3–540–28312–9.Google Scholar
 B Ganter. Two basic algorithms in concept analysis. Preprint 831, Technische (Hochschule Darmstadt, Germany, 1984).Google Scholar
 G Birkhoff. Lattice Theory First edition, Providence: American. Mathemathics Society (SpringlerVerlag BerlinHeidelberg, 2005), pp. 515–519. ISBN 10 3–540–28312–9.Google Scholar
 LA Zadeh, Fuzzy sets, information and control. 8:, 338–353 (1965).Google Scholar
 S Elloumi, J Jaam, A Hasnah, A Jaoua, I Nafkha, A multilevel conceptual data reduction approach based on the Lukasiewicz implication. Inf. Sci.163:, 253–262 (2004).View ArticleMathSciNetMATHGoogle Scholar
 J Riguet, Lattice Theory First edition, Relations binaires, fermetures et correspondances de Galois. Bull.Soc. Math. France. 78:, 114–155 (1948).MathSciNetGoogle Scholar
 R Belohlavek, Fuzzy Galois connections. Math. Logic Quart.45:, 497–504 (1999).View ArticleMathSciNetMATHGoogle Scholar
 R Belohlavek, Lattices of fixed points of Galois connections. Math.Logic Quart.47:, 111–116 (2001).View ArticleMathSciNetMATHGoogle Scholar
 A Frascella, Lattice Theory First edition, Fuzzy Galois connections under weak conditions, fuzzy sets and systems. 172:, 33–50 (2011).Google Scholar
 SN Srihari, H Arora, SH Cha, S Lee, Individuality of handwriting. J.Forensic Sci. 47(40), 1–17 (2002).Google Scholar
 HE Said, TN Tan, KD Baker, Personal identification based on handwriting. Pattern Recognit. 33(1), 149–160 (2000).View ArticleGoogle Scholar
 A Bensefia, T Paquet, L Heutte, A writer identification and verification system. Pattern Recognit. Lett.26(13), 2080–2092 (2005).View ArticleGoogle Scholar
 M Bulacu, L Schomaker, L Vuurpijl, in Seventh International Conference on Document Analysis and Recognition. Writer identification using edgebased directional features (IEEE, 2003), pp. 937–941.Google Scholar
 A Schlapbach, H Bunke, in 9th Int.Workshop on Frontiers in Handwriting Recognition. Using HMM based recognizers for writer identification and verification (IEEE, 2004), pp. 167–172.Google Scholar
 M Bulacu, L Schomaker, Textindependent writer identification and verification using textural and allographic features. IEEE Trans. Pattern Anal. Mach. Intell. 29(4), 701–717 (2007).View ArticleGoogle Scholar
 I Siddiqi, N Vincent, Text independent writer recognition using redundant writing patterns with contourbased orientation and curvature features. Pattern Recognit. 43(11), 3853–3865 (2010).View ArticleMATHGoogle Scholar
 U Pal, T Wakabayashi, F Kimura, in Ninth International conference on Document Analysis and Recognition ICDAR 07, 2. Handwritten numeral recognition of six popular scripts (IEEE Computer SocietyWashington, DC, USA, 2007), pp. 749–753. ISBN:0769528228.Google Scholar
 S Al Maadeed, F Ferjani, S Elloumi, Hassaine Ai, Jaoua A, in 2013 IEEE GCC Conference and exhibition, November 17–20, Doha, Qatar. Automatic handedness detection from offline handwriting (IEEE, 2013), pp. 119–124. ISBN: 9781479907229.Google Scholar
 UV Marti, R Messerli, H Bunke, in proceedings of Sixth International Conference on Document Analysis and Recognition. Writer identification using text line based features (IEEE Computer SocietyWashington, DC, USA, 2001), pp. 101–105.View ArticleGoogle Scholar
 EN Zois, V Anastassopoulos, Morphological waveform coding for writer identification. Pattern Recognit. 33:, 385–398 (2000).View ArticleGoogle Scholar
 S AlMa’adeed, W Ayouby, A Hassaine, J Aljaam. QUWI: An Arabic and English handwriting dataset for offline writer identification, International Conference on Frontiers in Handwriting Recognition (IEEE, 2012), pp. 746–751. ISBN: 9781467322621.Google Scholar
 N Otsu, A threshold selection method from graylevel histograms. Automatica. 11:, 285–296 (1975).View ArticleGoogle Scholar
 A Hassaine, S AlMaadeed, J AlJaam, A Jaoua, A Bouridane. The ICDAR2011 Arabic writer identification Contest, Proc. Eleventh International Conference on Document Analysis and Recognition,Beijing, China (IEEE, 2011).Google Scholar
 S Al Maadeed, A Hassaine, Automatic prediction of age, gender, and nationality in offline handwriting. EURASIP J Image Video Process.2014:, 10 (2014).View ArticleGoogle Scholar
 RO Duda, PE Hart, DG Stork, Pattern Classification, second edition (John Wiley & Sons Inc., New York, 2000).Google Scholar
 BV Dasarathy, Nearest Neighbor: Pattern Classification Techniques (IEEE Computer Society Press, New York, 1990).Google Scholar
 Y Yang, in Proc.17th Annual Intl. ACM SIGIR Conf. Research and Development in Information Retrieval, Dublin (Ireland). Expert network: effective and efficient learning from human decisions in text categorization and retrieval (ACM, 1994), pp. 13–22.Google Scholar
 L Breiman, Random forests. Mach. Learn.45(11), 5–32 (2001).View ArticleMATHGoogle Scholar
 G Salton, J Michael, An Introduction to Modern Information Retrieval (McGrawHill, New York, 1983).Google Scholar
 KDnuggets, Data integration, analytical ETL, data analysis, and reporting, rapid miner (2012). Software available at http://sourceforge.net/projects/rapidminer/.
 M Liwicki, A Schlapbach, H Bunke, Automatic gender detection using online and offline information. Anal. Appl. 14:, 87–92 (2011).MathSciNetGoogle Scholar