International Journal of Interactive Mobile Technologies (iJIM) – eISSN: 1865-7923 – Vol 17 No 08 (2023) Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine (SVM) Classifiers https://doi.org/10.3991/ijim.v17i08.39163 Mohammed Jawad Al_Dujaili1, Haider TH. Salim ALRikabi2(), Nisreen Khalil abed2, Ibtihal Razaq Niama ALRubeei2 1 Departement of Electronic and Communication, University of Kufa, Kufa, Iraq 2 Department of Electrical Engineering, Wasit University, Al Kut, Iraq hdhiyab@uowasit.edu.iq Abstract—In the realm of robotics and interactive systems, gender recogni- tion is a crucial problem. Considering the several uses it has in security, web search, human-computer interactions, etc., gender recognition from facial photos has garnered a lot of attention. The need to use and enhance gender recognition techniques is felt more strongly today due to a significant development in the design of facial recognition systems. Relatively speaking to other approaches, the progress gained in this area thus far is not exceptional. Thus, a novel method has been adopted in this study to improve accuracy in comparison to earlier research. To create the best rate of accuracy and efficiency in the suggested method of this research, we choose a minimal set of characteristics. Testing on the FERET and UTK-Face datasets reveals that our suggested algorithm has a lower degree of inaccuracy. In this article, the input image of the person's face is pre-processed to extract the right features from the face once the person's face has been recog- nized. Gender separation is achieved using Multi-class Support Vector Machine (SVM) Classifiers after features from normalized images have been extracted us- ing Histogram Oriented Gradient (HOG), Gabor Filters, and Speeded Up Robust Features (SURF), as well as their combination to select the most appropriate fea- ture from them as input for gender classification. As a feature reduction feature, the Principal Component Analysis (PCA) algorithm is also employed. Using the proposed approach, 98.75% gender recognition precision has been accomplished on the FERET database and a runtime performance of 0.4 Sec. on the UTK-Face database, 97.43% gender recognition accuracy has been accomplished and a runtime performance of 0.5 Sec. Keywords—gender recognition, Gabor, HOG, SURF, SVM 1 Introduction The face is a very important biometric component of a human being. It is through the face that valuable information such as race, identity, age, gender, and facial expres- sions are retrieved [1, 2]. In the topic of demography, Gender is an important feature of human beings. A human-computer interaction system can behave in a humanly manner iJIM ‒ Vol. 17, No. 08, 2023 113 https://doi.org/10.3991/ijim.v17i08.39163 mailto:hdhiyab@uowasit.edu.iq Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… when it can recognize the gender of the user. In the discussion of machine vision, au- tomatic gender recognition is a difficult problem with various applications. In human beings, gender recognition is often carried out through the human eyes. This recogni- tion is not easily done in machines. In the machine vision science, a lot of exertive methods have been done to develop and create algorithms and tools to appraise and recognize gender. If the person's gender can be identified prior to identification verifi- cation in the facial image problem, processing time is reduced by a half and face recog- nition speed is increased by double. In human-computer interaction, credit card secu- rity, advertising and psychology, surveillance, human-society, and criminal identifica- tion, automatic facial recognition and analysis serve a variety of purposes [3, 4]. The need to design systems that can recognize gender has risen and it continues to rise along with the notable advancement in the design of facial recognition systems. In machine vision, the automatic recognition of gender from facial images is because of challenging elements such as photo quality and light intensity, face position, the presence of emo- tions on the face, and whether the face is covered with a hat, scarf, and glasses [5, 6]. It is an arduous task, whereas gender recognition is an easy decision for human beings. Gender recognition can be utilized as a pre-processing step in face recognition, and since the studied space is cut in half (assumed to be equal for men and women), face recognition calculations can be performed twice as quickly. Pre-processing is the first general stage in the process of determining gender from a face photograph. The second stage is the Features of Extraction. Then the third stage is Categorization. These actions are displayed in Figure 1,2. The format of this article is such that, first, in the second part, an overview of the prior works is given, and then, in the third and fourth parts, respectively, the suggested approach and the outcomes of the application of various evaluation scenarios are described. Finally, we give a summary of the findings from the suggested method along with recommendations for the future. Fig. 1. Block Diagram of Gender Recognition from Face Images Fig. 2. Block Diagram of Gender Recognition from Face Images 114 http://www.i-jim.org Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… 2 Related works One of the most crucial and difficult challenges of the object recognition branch, which has various applications in human-computer interaction, human-society, psy- chology, and security issues, is the automatic face recognition and analysis. A binary classification problem, gender detection determines if the target image belongs to a man or a woman. Although humans can make this choice easily, machine recognition con- fronts a number of difficulties. Gender recognition can be employed as a pre-processing step in face recognition, and it doubles the speed of computations linked to face recog- nition since it cuts the examined space in half (assumes an equal number of men and women) [7-9]. According to the data used to identify gender, there have been several studies con- ducted in the previous 20 years that can be grouped into two primary categories. Tech- niques based on geometric traits fall into the first group, where they employ size infor- mation (distance between key facial features including the eyes, nose, lips, and chin) to determine gender [10, 11]. The methods based on appearance features fall into the sec- ond category. In these techniques, image pixels are transformed or subjected to mathe- matical operations [12, 13]. In comparison to other approaches, the development made in this area so far is insignificant. Many methods have been used for this purpose. Pre- trained Inception network, a neural technique based on the Face net model, is suggested in [14]. The suggested approach may be broken down into three steps: first, faces in the provided photos are recognized and clipped; second, the face images are sent to the neural network; and third, gender is classified with modified weights. The UTK Face dataset and 1x1, 3x3, and 5x5 filters have been used to train and test this network's model. Also, this article has a review of all machine learning techniques. In [15], they presented a technique for K-means clustering machine learning utilizing multi-Feature that, after improving the clarity and contrast of the images, uses a deep neural network to detect and cut the facial region. Then, using the FEI and SCIEN data sets, gender training and classification is carried out. Finally, features are extracted us- ing the SIFT descriptor. In [16], they employed discriminant error analysis and inde- pendent component analysis to enhance the gender classifier. They employed an eye detection system to apply this strategy to 500 images that were 96 x 64 in size. In [17], Baluja et al. used the Adabust method to identify gender in photos that were 12x12 in size. This network's design aims to create a quick method for gender detection in mas- sive databases. The etiquette algorithm, which was created by fusing multiple weak clauses, is utilized to speed up the process. The desired features are first retrieved using a feature extractor before being provided to the optimization method. For thumbnails, this approach is quick; but, if the network is given photos in big dimensions, the re- trieved features are too numerous, and the algorithm operates very slowly. A neural network technique utilizing the ICA algorithm is suggested in [18] to modify the weights. The Viola Jones method extracts facial features, which are subsequently se- lected using the NSGA-II rhythm algorithm. Finally, a hybrid neural network (ANN) with (ANN-ICA) performs gender categorization by recognizing facial features. Xu et al. performed face recognition by extracting facial outlines automatically and matching criterion points. With the use of the extracted features and the collection of iJIM ‒ Vol. 17, No. 08, 2023 115 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… two-dimensional photographs used to capture the face pattern, the differences between two faces are assessed. Four lines in front of the face and three lines on the side of the face are chosen using the feature-oriented orphans’ algorithm in this article. The accu- racy of the diagnosis increases with the number of selected components, as well as the calculation load [19]. Mian et al. used a region-based matching method to carry out automatic facial recog- nition. The forehead, eyes, nose, and chin sections of each photograph of a face in the gallery are separated into three groups using this technique. In order to prevent the im- pacts that can arise on faces with mustaches or beards, the diagnosis is only based on the forehead, eyes, and nose regions. The data collected indicate that the forehead and eyes are the key facial identification features [20]. Jofi et al in [21] focused on improv- ing gender recognition by concentrating on the regions around the eyes. In order to extract features, the author of [22] used PCA and an enhanced SIFT constant scale fea- ture transformation. In order to extract the features with a significant difference from photographs of the faces of women and men, this method was used to calculate the image matrix and choose the design of the input images through the clustering method. In a classifier based on fuzzy SVM, the author has employed a method based on gradual learning of the LVQ vector. The Gabor descriptor method is utilized in [23] combined with fuzzy classification-linear discriminant analysis to identify facial photographs from various age groups. 3 The proposed method The goal of this article is to discover the most effective and ideal technique for iden- tifying a person's gender from facial images. The standard procedure for identifying a person's gender from photos is reading the images from the image database and then extracting the facial image in accordance with the type of dataset. The characteristics are extracted after any necessary picture preprocessing. Following feature extraction, the size of the features is decreased to ensure good classification accuracy. Finally, the right classifier is chosen, and the photos are categorized or gendered. In this study, an attempt is made to use the facing image of the face. Features based on HOG, Gabor, and SURF as well as their combination are used for this purpose. Due to the demand for high-quality photos, several features are used, which increases the number of input data for classifiers and complicates categorization. Consequently, it is vital to appro- priately minimize the size of the features, and this reduction should be done in a way that improves the classification's accuracy. After the features have been extracted, the PCA approach will be used to minimize the problem's dimensions and increase the al- gorithm's classification accuracy. Finally, as an appropriate classifier, Multi-class Sup- port Vector Machine (SVM) is used. Figure 3 displays the proposed method's work- flow. 116 http://www.i-jim.org Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… Fig. 3. The Proposed Method for Face-Based Gender Recognition 3.1 Pre-processing Pre-processing is necessary for all of the photos in the image database, including normalization against brightness variations, scaling, and noise removal. The facial area is identified in this part utilizing image processing techniques. Each image in the dataset was subjected to the Viola-Jones technique, and the recognized faces were then scaled to a predetermined size of 128x128 pixels [24-27]. In Figures 4,5,6, the pre-processing procedures are displayed. Fig. 4. Image Preprocessing Phase [27] iJIM ‒ Vol. 17, No. 08, 2023 117 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… Fig. 5. Preprocessing on Sample from Images for FERET Database [28] Fig. 6. Preprocessing on Sample from images for UTK-Face Database [29] 3.2 Feature extraction Extraction of usable features from face photos is a crucial step in successful gender classification, much like in other applications of facial image analysis. Even the finest classifiers run the risk of falling short of achieving high recognition accuracy if inade- quate features are used. Facets of the image are retrieved after the pre-processing stage. The retrieved features are often divided into two groups: global features, or features based on appearance, and local features, or features based on geometry [30-33]. Instead of extracting characteristics from individual face points, the appearance-based method extracts features from the entire face. These techniques rely on various adjustments and adjustments made to the image's pixels. Face characteristics like the nose and eyes are retrieved using the geometry-based technique. The relationships between the face points are typically used to extract the constant geometric properties, such as scale, rotation, distance, and angle. The characteristics provide a trained classifier access and represent the human face. The characteristics of the Histogram Oriented Gradient (HOG), Gabor Filters, and SURF are used in this article and combined as input for gender classification. Histogram Oriented Gradient (HOG). Histogram Oriented Gradient is based on computing the gradient orientation histograms for each cell by dividing an image into 118 http://www.i-jim.org Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… smaller cells. Following that, the histograms are combined into a single feature vector that describes the overall structure of the image. HOG is especially helpful for locating items with a distinct shape, like people, animals, or cars. It is suitable for use in practical applications since it is resistant to variations in illumination and viewpoint [34] HOG- based facial image gender detection is a prevalent issue in computer vision and machine learning. Typically, to distinguish between male and female faces, a classifier, such as an SVM, is trained by first extracting HOG characteristics from face pictures. The pho- tos are initially pre-processed to identify and align the faces in order to extract HOG features from the face images. Algorithms for face detection and facial landmark de- tection are commonly used for this. A single feature vector is then created by concate- nating the HOG features computed for each face region [35, 36]. A classifier can be trained to distinguish between male and female faces using a dataset of labeled face images after the HOG characteristics have been extracted. By generating a new face image's HOG characteristics and utilizing the trained classifier to make a prediction, one can then use the trained classifier to predict the gender of the image. It's crucial to remember that the caliber of the training data and the classifier selected will determine how accurate a gender detection system based on HOG characteristics is. Furthermore, occlusions, illumination, and changes in facial expression may have an impact on it [37]. We perform the following steps to calculate the gradient histogram: To extract the picture gradient in the x and y directions, the image is first filtered using Sobel kernels in the x and y directions: 𝐺𝐺𝑥𝑥 = 𝐼𝐼 ∗ 𝐷𝐷𝑥𝑥 (1) 𝐺𝐺𝑦𝑦 = 𝐼𝐼 ∗ 𝐷𝐷𝑦𝑦 (2) I: is the original image, D x and D y are the Sobel kernels in the x, y direction, G x, G y are the image gradient in the y, x direction, and the sign*indicates the convolution operation. The gradient direction's size in each pixel is then determined as follows : |𝐺𝐺(𝑖𝑖, 𝑗𝑗)| = �(𝐺𝐺𝑥𝑥(𝑖𝑖, 𝑗𝑗))2 + (𝐺𝐺𝑦𝑦(𝑖𝑖, 𝑗𝑗))2 (3) 𝜃𝜃𝐺𝐺(𝑖𝑖, 𝑗𝑗) = 𝑡𝑡𝑡𝑡𝑡𝑡−1 � 𝐺𝐺𝑥𝑥(𝑖𝑖,𝑗𝑗) 𝐺𝐺𝑦𝑦(𝑖𝑖,𝑗𝑗) � (4) That |G| The number of rows and columns in the image are represented by j and I respectively, and the gradient size θ G indicates the gradient's direction. For color im- ages, the gradient is determined for each color channel independently, and the gradient vector for each pixel is chosen based on its biggest value. We initially restrict the gra- dient angle to the range of 0-180 degrees to calculate the gradient histogram as follows: 𝜃𝜃𝐺𝐺′ = � 𝜃𝜃𝐺𝐺 , 0 ≤ 𝜃𝜃𝐺𝐺 < 180 𝜃𝜃𝐺𝐺 − 180 , 180 ≤ 𝜃𝜃𝐺𝐺 < 360 (5) The range from 0 to 180 degrees is then divided into n equal distances, each of which represents a histogram channel and indicates the number of gradient directions or his- togram intervals. iJIM ‒ Vol. 17, No. 08, 2023 119 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… Gabor filters. Gabor filters may encode the local frequency and orientation infor- mation of an image, making them ideal for feature extraction from images [38]. As a result, they can be used to identify edges, textures, and other components in an image. Gabor filters are frequently used in computer vision to extract characteristics from im- ages for tasks including object detection, texture categorization, and face recognition. By convolving the image with the filter, the filters are applied to an image, producing a filtered image that emphasizes particular aspects of the original image. The attributes that are pertinent for the task at hand can subsequently be extracted from the filtered image by additional processing. The frequency and orientation of gabor filters are often adjustable, and these properties can be changed to match the important aspects of an image. They may thus be tailored to various image kinds and activities, making them a versatile tool for feature extraction. A computer vision task called "gender detection from face images using Gabor filters" entails first extracting features from the faces in the images using the Gabor filters, and then categorizing the faces according to their gender. The first step in applying Gabor filters to do gender recognition from face pho- tographs is to pre-process the images to find and align the faces. Algorithms for face detection and facial landmark detection are commonly used for this. Gabor wavelet and Gabor filter are directly connected. A sine wave and a Gaussian wave are combined to create a Gabor wavelet. Since this filter operates in the frequency domain, it must first be created. Once it has been created, the images must also be converted to the frequency domain before the filter can be applied to the image [39]. Equation (6) is used to derive Gabor filter: 𝑔𝑔(𝑥𝑥, 𝑦𝑦, λ, σ, γ, θ, φ) = 𝑔𝑔𝑅𝑅(𝑥𝑥, 𝑦𝑦, λ, σ, γ, θ, φ) + 𝑗𝑗𝑔𝑔1(𝑥𝑥, 𝑦𝑦, λ, σ, γ, θ, φ) (6) In formula 1, there are gR analytical and g1 imaginary parameters of the Gabor filter, which are calculated from formulas (7,8). 𝑔𝑔𝑅𝑅(𝑥𝑥, 𝑦𝑦, λ, σ, γ, θ, φ) = 𝛾𝛾 2𝜋𝜋𝜋𝜋2 𝑒𝑒� 𝑥𝑥𝑟𝑟 2+𝑦𝑦2𝑦𝑦𝑟𝑟 2 2𝜎𝜎2 � cos �2𝜋𝜋 𝑥𝑥𝑟𝑟 𝜆𝜆 + 𝜑𝜑� (7) 𝑔𝑔1(𝑥𝑥, 𝑦𝑦, λ, σ, γ, θ, φ) = 𝛾𝛾 2𝜋𝜋𝜋𝜋2 𝑒𝑒( 𝑥𝑥𝑟𝑟 2+𝑦𝑦2𝑦𝑦𝑟𝑟 2 2𝜎𝜎2 ) sin �2𝜋𝜋 𝑥𝑥𝑟𝑟 𝜆𝜆 + 𝜑𝜑� (8) In formulas (7,8), parameters xr and yr are calculated using formulas (9,10). 𝑥𝑥𝑟𝑟 = 𝑥𝑥𝑥𝑥𝑥𝑥𝑥𝑥𝜃𝜃 + 𝑦𝑦𝑥𝑥𝑖𝑖𝑡𝑡𝜃𝜃 (9) 𝑦𝑦𝑟𝑟 = 𝑥𝑥𝑥𝑥𝑖𝑖𝑡𝑡𝜃𝜃 + 𝑦𝑦𝑥𝑥𝑥𝑥𝑥𝑥𝜃𝜃 (10) The coordinates of a point in the image are represented by the parameters x and y. The Gabor filter's rotation angle is displayed by the parameter. The phase offset is the parameter φ. This parameter displays the Gabor filter's symmetry. The Gabor filter is either odd or even for φ values of 0 and 180, and it is either asymmetric or odd for φ values of 90 and -90. The value of the variance associated with the Gaussian function is specified by parameter σ. The spatial graph rate is parameter y. The supported range has a circular form I y f is equal to 1 and rotates to ellipses in the direction if it is less than 1. The sine wave frequency's wavelength is represented by the parameter λ. 120 http://www.i-jim.org Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… Speeded Up Robust Features (SURF). A computer vision method called SURF is used to find and describe small details in images. A SIFT algorithm known as SURF is used to locate objects of interest in images that are independent of scale and orientation. The SURF algorithm is a quicker and more reliable variant of the SIFT method [40]. While using SURF, an image's stable key points are first found, and then their de- scriptors are computed. In order to match key points between images, descriptors— vectors that characterize the local image material surrounding the key point—are used. In many computer vision applications, including object detection, picture matching, and tracking, SURF is employed. In real-time processing-intensive applications like aug- mented reality and video surveillance, it is also helpful. A computer vision method called SURF is used to identify and describe features in photographs, especially in dig- ital portraits of people. SURF is a valuable technique for gender recognition since the traits it picks up can be utilized to represent a picture and set it apart from other images [41]. The following steps are usually observed to use HOG, Gabor Filters and SURF technique for gender recognition: 1. Preprocessing: Input face images are preprocessed in order to enhance contrast, re- move noise, and to align the faces to a standard pose. 2. Feature extraction: To extract the features from the preprocessed face images, HOG, Gabor Filters and SURF is used. These features are stored appropriately as a set of points with related descriptions. 3. 3-Dimensionality reduction: To decrease the risk of over fitting, the feature points and descriptions are usually reduced to speed up the classification process in dimen- sionality. 4. 4-Model training: The training of a machine learning model is carried out on the reduced feature representation through the use of labeled training data. 5. 5-Model evaluation: The evaluation of the trained model is carried out on a different validation set so that its accuracy can be measured. 6. Deployment: In a gender recognition system, the trained model is positioned where it can be used to arrange new face photographs into male or female categories. 3.3 Combining features and selection The process of joining or merging different attributes of a data set into one feature representation refers to combining features. This is a common step in feature engineer- ing, where the goal is to create powerful and meaningful depiction of data that can be used in machine learning algorithms. Various methods for combining features include: 1. Concatenation: This is basically connecting several features into a distinct vector, ensuing in a higher-dimensional feature representation. 2. Feature Crosses: Merging the values of two or more existing features in a non-linear manner in order to create new features. 3. Averaging: combining the average of multiple features into a one feature. The chosen feature combining method will be dependent on the characteristics of the data and the specific problem being solved. It is crucial to thoroughly consider the iJIM ‒ Vol. 17, No. 08, 2023 121 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… effect of merging features on the performance of the machine learning algorithm, and to experiment using diverse methods to determine the best method [30]. We made use of all the above-mentioned combining methods in this article, and the Feature Crosses method emerged to be the best method. Creating new features by merging the values of two or more existing features in a non-linear way is defined as feature crosses. The purpose of feature crosses is to record interactions between features that the individual features do not record on their own. 3.4 Reducing the dimensions of the feature vector PCA is a technique statistically used to decrease the dimensionality of data while maintaining a lot of information. It is a globally used method in data analysis and it is useful particularly for envisioning high-dimensional data, such as images in a lower- dimensional space [42]. The primary goal of PCA is to identify a new collection of orthogonal axes, known as principal components that will effectively capture the most significant information in the data. PCA is frequently used in pre-processing machine learning algorithms, as dimensionality reduction of data can foster the development and performance of the algorithm and diminish over fitting. It can be utilized for compress- ing data as well, because the first few key components naturally record most of the information in the data, and the rest of it can be gotten rid of. PCA is an unsupervised method, which means it is based on the statistical characteristics of the data rather than any labeled information about the data. PCA has some drawbacks despite its ease of use and widespread application, including its sensitivity to outliers and the fact that it can only identify linear correlations in the data. It is still a powerful and widespread tool for data analysis [43, 44]. The following are the steps in this method: The first step involves transforming the input image which is in the form of a two-dimensional ma- trix, into a one-dimensional vector and create the feature matrix before combining these vectors. The input image, which is a two-dimensional matrix, must first be transformed into a one-dimensional vector in the first phase before these vectors can be combined to create the feature matrix. The feature matrix is obtained as n*N, where n=p*q, if the total number of our photos is N and the size of each image is n=p*q. 𝑋𝑋 = (𝑥𝑥1, 𝑥𝑥2, … … . . 𝑥𝑥𝑖𝑖, … … 𝑥𝑥𝑁𝑁) (11) The feature vector for each image with n dimensions in relation (11) X, the feature matrix xi, is n*N, where n=p*q The average will then be calculated. In this stage, we first determine the average of each category of data, and then we deduct each category's data from its average until the data is converted to zero. 𝜇𝜇 = 1 𝑁𝑁 ∑ 𝑥𝑥𝑖𝑖 𝑁𝑁𝑖𝑖=1 (12) In the third phase, the covariance matrix is calculated, and the eigenvalues and ei- genvectors are then extracted from this matrix. 𝑆𝑆𝑇𝑇= ∑ (𝑥𝑥𝑖𝑖 − 𝜇𝜇)(𝑥𝑥𝑖𝑖 − 𝜇𝜇)𝑇𝑇𝑁𝑁𝑖𝑖=1 (13) 122 http://www.i-jim.org Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… The idea of decreasing data dimensions is presented in the final step. Based on their special values, we order the special vectors acquired in the previous phase from large to small. The data elements are then arranged in order of importance, starting with the most important. In this case, we can eliminate the less significant data if we wish to lower the dimensions of the data. Of course, by doing this, we lose some of our infor- mation from the special space, which is the river of space created by these special vec- tors. The following results can be achieved by implementing this methodology to the data: 1. Orthogonalization of input vector components. 2. Sorting and ranking the major components. 3. Eliminating components with little alterations PCA is displayed for two-dimensional data in Figure 7. Fig. 7. Distribution of the data before and after processed to PCA 3.5 Classification The classification is the final step in gender recognition. The level of classification accuracy is determined by the features extracted in the feature extraction step. The clas- sification of faces is carried out by the classifier which categorizes the face images into two classes: female and male. As far as this step is concerned, a wide range of classifiers have been used overtime such as, K_NN [45], ANN [5], and SVM [46]. Gender recog- nition of humans has been achieved in this study by using Multi-class SVM Classifiers. This step involves training the SVM classifier so that it will be able to differentiate between multiple classes like female and male. Basically, a different binary classifier is trained for each pair of classes, and then the classifiers are combined to produce a multiclass classifier. This can be achieved using numerous methods such as error-cor- recting codes and one-vs-all. The latter is a simple technique which involves training a separate binary classifier for each class, with one class being positive and the others being negative. Consequently, the classifier which produces the highest output is then chosen as the prediction for the input. It is also a more complex approach that involves training a separate binary classifier for each pair of classes. The classifier is fed with input sample, and the outputs of the classifier are merged to determine the final predic- tion. On the other hand, the error-correcting output codes is a more advanced approach whereby, a set of binary classifiers are trained, and then the classifiers’ outputs are used iJIM ‒ Vol. 17, No. 08, 2023 123 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… to make a final prediction. The main idea of this approach is to make use of the classi- fiers’ outputs to correct any errors made by individual classifiers. For the recognition of human gender, a dataset of labeled face image can be used to train a multi-class SVM and to predict the gender of a new face image. A technique is selected based on the dataset’s size as well as the desired accuracy of the classifier [47]. Figure 8 presents an image of a dataset belonging to two classes that are selected by the SVM method as the best hyper surface for their separation. Fig. 8. SVM classification Primarily, a linear decision boundary is presented as follows: 𝑊𝑊. 𝑋𝑋 + 𝑏𝑏 = 0 (14) where the multiplication sign W denotes the normal vector, which is positioned per- pendicularly to the super plane. In order to maximize the distance between the parallel super planes separating the data, W,b are selected. Equations (15, 16) are used to described the hyper planes. 𝑊𝑊. 𝑋𝑋 + 𝑏𝑏 = 1 (15) 𝑊𝑊. 𝑋𝑋 + 𝑏𝑏 = −1 (16) In a situation whereby the training data can be separated linearly, two hyper planes on the edge of the points re considered so that they do not have common points, and afterwards, their distance is maximizing. For the distance to maximized, the soft W must be minimized. Points are prevented from getting into the margin by making addi- tions of the following conditions: for each i, one of the following conditions must be satisfied, relation (17,18). 𝑊𝑊. 𝑋𝑋𝑖𝑖 − 𝑏𝑏 ≥ 1 for Xi (17) 𝑊𝑊. 𝑋𝑋𝑖𝑖 − 𝑏𝑏 ≤ −1 for Xi (18) 124 http://www.i-jim.org Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… The first condition is for the first class 𝑋𝑋𝑖𝑖 while the second condition is for the second class 𝑋𝑋𝑖𝑖. The following expression represents the combination of the two conditions: 𝑦𝑦𝑖𝑖(𝑤𝑤. 𝑥𝑥𝑖𝑖 + 𝑏𝑏) = 1 (19) 𝑦𝑦𝑖𝑖(𝑤𝑤. 𝑥𝑥𝑖𝑖 + 𝑏𝑏) > 1 (20) The following optimization problem must be solved as a way of calculating the op- timal decision boundary 𝑚𝑚𝑡𝑡𝑥𝑥𝑤𝑤,𝑏𝑏 𝑚𝑚𝑖𝑖𝑡𝑡𝑖𝑖=1,…,𝑙𝑙 = �𝑦𝑦𝑖𝑖 𝑤𝑤.𝑥𝑥𝑖𝑖+𝑏𝑏 |𝑤𝑤| � (21) Base on Eq (21) and other mathematical calculations, the relationship above can be converted into the relationship below: 𝑚𝑚𝑖𝑖𝑡𝑡𝑤𝑤,𝑏𝑏 1 2 |𝑤𝑤|2, 𝑦𝑦𝑖𝑖(𝑤𝑤. 𝑥𝑥𝑖𝑖 + 𝑏𝑏) − 1 > 0 𝑖𝑖 = 1, … . , 𝐿𝐿 (22) Solving the optimization problem (22) is a challenging operation. However, it can be simplified through the use of Lagrange's indefinite coefficients method; this optimi- zation problem can be transformed into the following form where λ𝑖𝑖 denotes Lagrange's coefficients. 𝑚𝑚𝑡𝑡𝑥𝑥λ1,..λ𝐿𝐿 = �− 1 2 ∑ ∑ λ𝑖𝑖𝑦𝑦𝑖𝑖𝐿𝐿𝑗𝑗=1 𝐿𝐿 𝑖𝑖=1 �𝑥𝑥𝑖𝑖, 𝑥𝑥𝑗𝑗�λ𝑗𝑗𝑦𝑦𝑗𝑗 + ∑ λ𝑖𝑖 𝐿𝐿 𝑖𝑖=1 � λ𝑖𝑖 ≥ 0 𝑖𝑖 = 1, … , 𝐿𝐿 ∑ λ𝑖𝑖𝐿𝐿𝑖𝑖=1 𝑦𝑦𝑗𝑗 = 0 (23) Once the optimization problem has been solved, the equation below is used to cal- culate the Lagrange coefficients. 𝑤𝑤 = ∑ λ𝑖𝑖𝐿𝐿𝑖𝑖=1 𝑦𝑦𝑗𝑗𝑥𝑥𝑖𝑖 (24) The support vectors are greater than zero, and λ𝑖𝑖 will be zero at other points. Thus, giving that λ𝑖𝑖 equals 6 and zero, corresponding to 𝑥𝑥𝑖𝑖 which are not support vectors, only few training points which are support vectors are required for the derivation of the de- cision boundary, and they are all unnecessary. Consequently, just few training points will be required to classify hyper spectral images using support vector. Once w is de- rived using the following relationship, the calculation of b is done for a wide range of support vectors, and the all the b’s derived are averaged in order to obtain the final b. λ𝑖𝑖[𝑦𝑦𝑖𝑖(𝑤𝑤. 𝑥𝑥𝑖𝑖 + 𝑏𝑏) − 1] = 0 𝑖𝑖 = 1, … . , 𝐿𝐿 (25) Eq (26) is used to obtain the final classifier. 𝑓𝑓(𝑋𝑋, 𝑊𝑊, 𝑏𝑏) = 𝑥𝑥𝑔𝑔𝑡𝑡(𝑊𝑊. 𝑋𝑋 + 𝑏𝑏) (26) In the case of multi-class, the aim is to obtain the hyper plane that separates the classes and to also maximize the margin between the classes. iJIM ‒ Vol. 17, No. 08, 2023 125 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… Due to the ability of the multi-class SVM to handle large datasets and the ease with which it is implemented, it was used in the proposed technique. It demonstrated supe- rior performance as presented in the results section. The primary function for a multi-class SVM can represented as given below: 𝐿𝐿(𝑤𝑤, 𝑏𝑏) = 1 2 �|𝑤𝑤|� 2 + 𝐶𝐶 𝛴𝛴{𝑖𝑖=1} 𝑛𝑛 𝛴𝛴{𝑗𝑗=1} 𝐶𝐶 [𝑦𝑦𝑖𝑖 = 𝑗𝑗]𝑚𝑚𝑡𝑡𝑥𝑥 �0, 1 − 𝑦𝑦𝑗𝑗�𝑤𝑤. 𝑥𝑥𝑖𝑖 + 𝑏𝑏𝑗𝑗�� (27) where w denotes weight vector, b is bias vector, 𝑥𝑥𝑖𝑖 ith feature vector, 𝑦𝑦𝑖𝑖class label for the ith feature vector, 𝑦𝑦𝑗𝑗class label for class j, C a regularization parameter that trades off the margin size and the misclassification error, n number of samples for train- ing. The use of a quadratic programming algorithm can be employed in solving the problem. Using this solution, biases and weights for each class can be derived and em- ployed in making predictions for new data points. In the one-vs-all approach, the multi- class problem is minimized to multiple binary classification problems, and a separate hyper plane is derived for each class. Here, the objective function is represented as follows: 𝐿𝐿(𝑤𝑤, 𝑏𝑏) = 1 2 �|𝑤𝑤|� 2 + 𝐶𝐶 𝛴𝛴{𝑖𝑖=1} 𝑛𝑛 𝑚𝑚𝑡𝑡𝑥𝑥�0, 1 − 𝑦𝑦𝑖𝑖(𝑤𝑤. 𝑥𝑥𝑖𝑖 + 𝑏𝑏)� (28) where w is the weight vector, b denotes bias value, 𝑥𝑥𝑖𝑖 ith feature vector, 𝑦𝑦𝑖𝑖 class la- bel for the ith feature vector, C is a regularization parameter through which a tradeoff between the misclassification error, margin size, and n number of training samples is achieved. 4 Experimental results and analysis Generally, the steps involved in writing this paper are highlighted as follows: 1. Data Preparation 2. Modification of raw data to be fed to the network 3. Creation of an appropriate network 4. Training of the network. Generally, the process involves taking the sample as raw data in the first step. Sub- sequently, the use of the pre-processing function is employed in eliminating the initial parts of the raw data. An example of the pre-processing function applied is normalizing the data against changes in brightness, scale, and elimination of noise. The training and evaluation of the method presented in this study are carried out using FERET and UTK- Face databases. More than 14,000 images of 1,000 individuals obtained over a period of many years are contained in the database used in this study. The images were cap- tured under regulated conditions, and are characterized by numerous poses, facial ex- pressions, and lighting conditions. FERET is one of the benchmark datasets that is ex- tensively used in the field of facial recognition, and its application has been made in a wide range of researches to assess and compare the performance of different facial recognition algorithms. Researchers in the field of facial recognition are presented with 126 http://www.i-jim.org Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… a challenging test set in the dataset. The results derived from experiments using FERET database are extensively cited [23]. On the other hand, the UTK-Face is a large dataset containing more than 20,000 face images captured from people of different genders, ages, and ethnicities. The images contained in the database were captured and collated by the University of Tennessee, Knoxville. It is a robust dataset that contains a variety of facial images for machine learning and computer vision research. Extensive use of the UTK-Face dataset is employed research and development in different applications, including estimation of age, recognition of age, and analysis of facial features. The da- taset contains annotations for the facial attributes, like ethnicity, gender, and age, mak- ing it possible to train and assess machine learning models for the aforementioned tasks [24]. Subsequent to the reading of images from the dataset and application of image preprocessing, features extraction is performed, and then they are combined together so that the most suitable feature is selected. After the selection, features reduction is carried out with the aim of achieving high classification accuracy. Lastly, the selection of the most suitable classifier is made and used to classify the images of determine gender. In this study, this achieved using the face image of individuals. To accomplish this purpose, features based on Gabor, HOG, and SURF are combined and used. The use of multiple features because of the need to obtain high-quality images, results in an increase in the volume of input data for classifiers, thereby making the process of clas- sification challenging. Thus, it becomes of important for the reduction of features’ di- mensions to be done in the most appropriate manner so that the classification accuracy can be improved. For this reason, upon completion of the process of features extraction for the reduction of the problems’ dimensions, the PCA algorithm is used to enhance the accuracy of classification. Finally, Multi-class SVM is used as a suitable classifier. The data obtained from the FERET and UTK-Face databases is used to test the im- plemented model. These tests also involved the evaluation of the performance of the proposed model from different perspectives, and the results have been presented. Based on the experimental tests, the proposed model has demonstrated high level of accuracy in terms of recognizing the gender of people using images of their faces, and can be employed as an efficient tool in real applications. The performance evaluation of the proposed method was done based on some parameters including root mean square error. This parameter shows the mean square difference between the values predicted by the proposed method and the real age of people. The mean square of error represents the square of the prediction error of the classification algorithm, and it is derived through the following equation: 𝑅𝑅𝑅𝑅𝑆𝑆𝑅𝑅 = �1 𝑁𝑁 ∑ (𝑝𝑝𝑖𝑖 − 𝑥𝑥𝑖𝑖)2𝑁𝑁𝑖𝑖=1 (29) In the relationship presented above, N is the number of test samples, 𝑝𝑝𝑖𝑖 denotes the value predicted for the age of the individual in the test sample i, and 𝑥𝑥𝑖𝑖 represents the value of the real age in this sample. This aimed at deriving the lowest mean squared error in prediction. Normal error percentage: this parameter is used calculate the error percentage of the prediction model, and this can be derived by computing the ratio between the error’s iJIM ‒ Vol. 17, No. 08, 2023 127 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… average absolute value and the series of changes of the target values. This is calculated using the formulae below: 𝑁𝑁𝑥𝑥𝑁𝑁𝑚𝑚𝑡𝑡𝑁𝑁𝑖𝑖𝑁𝑁𝑒𝑒𝑁𝑁𝑒𝑒𝑟𝑟𝑟𝑟𝑒𝑒𝑟𝑟 = 100 × 𝑀𝑀𝑀𝑀𝑀𝑀 𝑚𝑚𝑚𝑚𝑥𝑥(𝑠𝑠)−𝑚𝑚𝑖𝑖𝑛𝑛 (𝑠𝑠) (30) In the relationship above, s represents the actual values of the target in the test sam- ples, and min and max represent the calculation functions of the maximum and mini- mum, respectively. The error’s average absolute value: this parameter is used to determine the difference between the average predicted values for the people’s age and the actual state. The equation below is used to derive the average absolute value of the error: 𝑅𝑅𝑀𝑀𝑅𝑅 = 1 𝑛𝑛 ∑ |𝑦𝑦𝑖𝑖 − 𝑦𝑦𝚤𝚤�|𝑛𝑛𝑖𝑖=1 (31) The aim of using this parameter is obtain the minimum average absolute value of the error for prediction. In the approach proposed in this work, equation 18 below was used to calculate the average absolute error: Where n denotes the number of data points, 𝑦𝑦𝑖𝑖 is the actual value of the i-th data point. 𝑦𝑦𝚤𝚤� is the predicted value of the i-th data point. ∑ |𝑦𝑦𝑖𝑖 − 𝑦𝑦𝚤𝚤�|𝑛𝑛𝑖𝑖=1 represents the sum of the absolute differences between the actual and predicted values over all data points. Tables 1 and 2 present the results of simulation performed using the FERET database. It can be seen from the tables that the proposed technique was evaluated based on seven features, out of which three are individual (Ga- bor Filters, HOG and SURF), and the remaining four are combined (HOG+ Gabor, HOG+ SURF, Gabor+ SURF and HOG+ Gabor+ SURF). The results obtained from the simulation are significantly influenced by the selection of the most suitable feature. Also, the accuracy of the system increases when the most appropriate feature is chosen. Rather than using individual features, combined features were used, and this resulted in better system performance. However, the use of the combined features caused an increase in the running time of the algorithm as seen in Tables 1 and 2. This can be attributed to the increased size of features resulting from combining them, and the use of PCA algorithm is employed in addressing the issue of increased running time. The PCA algorithm achieves this by deleting features that do not affect the results of simu- lation, thereby increasing the speed of the algorithm as well as the accuracy of the re- sults. It can be clearly seen from Table 2 that PCA has an effect on the proposed tech- nique, as the results presented in Table 1 show a higher running time and lower accu- racy as compared to the results on Table 2. The results presented in Table 1 were ob- tained before the application of the PCA algorithm, while that on Table 2 were obtained after the use of PCA algorithm was employed. From the Table 2 it can also be seen that when HOG+ Gabor+ SURF features are combined, higher accuracy and lesser running time are achieved as compared to using individual features. Tables 3 and 4 present the results of simulation for UTK-Face database. It can be seen from Tables 1 and 2 that the simulation results and performance of the system are significantly influenced by the 128 http://www.i-jim.org Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… combination of features and selection of the most appropriate features. The combina- tion of features also improves the accuracy of the system as compared to when individ- ual features are used. More so, combining the HOG+ Gabor+ SURF features reduces the running time of the system. Table 1. Percentage accuracy and a runtime with FERET database for all features without PCA algorithm Features HOG Gabor SURF HOG+ Gabor HOG+ SURF Gabor+ SURF HOG+ Gabor+ SURF Accu %. 90.77 91.63 90.36 94.28 94.45 95.67 96.98 Time Sec 0.6 0.6 0.7 0.8 0.8 0.8 0.8 Table 2. Percentage Accuracy and a runtime with FERET Database for All Features with PCA Algorithm Features HOG Gabor SURF HOG+ Gabor HOG+ SURF Gabor+ SURF HOG+ Gabor+ SURF Accu %. 93.82 95.43 95.39 96.41 96.1 97.67 98.75 Time Sec 0.2 0.2 0.2 0.3 0.3 0.3 0.4 Table 3. Percentage accuracy and a runtime with UTK-Face database for all features without PCA algorithm Features HOG Gabor SURF HOG+ Gabor HOG+ SURF Gabor+ SURF HOG+ Gabor+ SURF Accu %. 91.42 91.66 91.31 93.25 95.43 95.39 96.46 Time Sec 0.5 0.6 0.6 0.7 0.8 0.8 0.9 Table 4. Percentage Accuracy and A Runtime with UTK-Face Database for All Features with PCA Algorithm Features HOG Gabor SURF HOG+ Gabor HOG+ SURF Gabor+ SURF HOG+ Gabor+ SURF Accu %. 94.17 94.88 95.76 97.27 96.91 97.22 97.43 Time Sec 0.3 0.3 0.2 0.5 0.4 0.5 0.5 5 Conclusion This research was aimed at finding an efficient and optimal technique for the recog- nition gender using facial images. This was achieved by determining the most critical features of people’s faces that should be considered, and how the features should be applied and utilized to achieve the best performance. Features based on Gabor, HOG, and SURF, and a combination of the features are used in selecting the appropriate fea- ture from them as input for gender classification. The use of multiple features due to the need for high-quality images results in increased volume of input data for classifiers, iJIM ‒ Vol. 17, No. 08, 2023 129 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… thereby making the process of classifying the features very challenging. Thus, it is im- portant that the features’ dimensions be reduced in a manner that results in improved classification accuracy. For this reason, the use of the PCA is employed after the fea- tures have been extracted for reduction of dimensions of the problem and for the im- provement of classification accuracy. Lastly, efficient and suitable classifiers by Multi- class SVM were used in this work, with an accuracy of 98.75% achieved by the pro- posed technique in terms of recognition of gender on the facial images obtained from the FERET database and a runtime execution of 0.4 Sec., with gender recognition ac- curacy of 97.43% achieved for the facial images obtained from the UTK-Face database and a runtime execution 0.5 Sec. 6 References [1] W. Wu, P. Protopapas, Z. Yang, and P. Michalatos, "Gender classification and bias mitiga- tion in facial images," in 12th acm conference on web science, 2020, pp. 106-114. https://doi.org/10.1145/3394231.3397900 [2] Y. Lin and H. Xie, "Face gender recognition based on face recognition feature vectors," in 2020 IEEE 3rd International conference on information systems and computer aided education (ICISCAE), 2020: IEEE, pp. 162-166. https://doi.org/10.1109/ICISCAE51034. 2020.9236905 [3] T. A. Sumi, M. S. Hossain, R. U. Islam, and K. Andersson, "Human gender detection from facial images using convolution neural network," in Applied Intelligence and Informatics: First International Conference, AII 2021, Nottingham, UK, July 30–31, 2021, Proceedings 1, 2021: Springer, pp. 188-203. https://doi.org/10.1007/978-3-030-82269-9_15 [4] H. Al-ogaili and A. M. Shadhar, "the Finger Vein Recognition Using Deep Learning Technique," Wasit Journal of Computer and Mathematics Sciences, vol. 1, no. 2, pp. 1-11, 2022. [5] J. Rwigema, J. Mfitumukiza, and K. Tae-Yong, "A hybrid approach of neural networks for age and gender classification through decision fusion," Biomedical Signal Processing and Control, vol. 66, p. 102459, 2021. https://doi.org/10.1016/j.bspc.2021.102459 [6] R. ALairaji, and H. Salim, "Abnormal Behavior Detection of Students in the Examination Hall From Surveillance Videos," in Advanced Computational Paradigms and Hybrid Intelligent Computing, vol. 1373: Springer Singapore, 2022, pp. 113-125. https://doi.org/ 10.1007/978-981-16-4369-9_12 [7] S. Kumar, S. Singh, J. Kumar, and K. Prasad, "Age and gender classification using Seg-Net based architecture and machine learning," Multimedia Tools and Applications, vol. 81, no. 29, pp. 42285-42308, 2022. https://doi.org/10.1007/s11042-021-11499-3 [8] S. Fekri-Ershad, "Gender classification in human face images for smart phone applications based on local texture information and evaluated Kullback-Leibler divergence," Traitement du Signal, vol. 36, no. 6, pp. 507-514, 2019. https://doi.org/10.18280/ts.360605 [9] H. Sabah, "A Detection of Deep Fake in Face Images Using Deep Learning," Wasit Journal of Computer and Mathematics Sciences, vol. 1, no. 4, pp. 94-111, 2022. [10] V. K. Verma, S. Srivastava, T. Jain, and A. Jain, "Local invariant feature-based gender recognition from facial images," in Soft Computing for Problem Solving: SocProS 2017, Volume 2, 2019: Springer, pp. 869-878. https://doi.org/10.1007/978-981-13-1595-4_69 [11] R. A. Azeez, M. K. Abdul-Hussein, M. S. Mahdi, and H. T. S. ALRikabi, "Design a system for an approved video copyright over cloud based on biometric iris and random walk 130 http://www.i-jim.org https://doi.org/10.1145/3394231.3397900 https://doi.org/10.1109/ICISCAE51034.2020.9236905 https://doi.org/10.1109/ICISCAE51034.2020.9236905 https://doi.org/10.1007/978-3-030-82269-9_15 https://doi.org/10.1016/j.bspc.2021.102459 https://doi.org/10.1007/978-981-16-4369-9_12 https://doi.org/10.1007/978-981-16-4369-9_12 https://doi.org/10.1007/s11042-021-11499-3 https://doi.org/10.18280/ts.360605 https://doi.org/10.1007/978-981-13-1595-4_69 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… generator using watermark technique," Periodicals of Engineering Natural Sciences, vol. 10, no. 1, pp. 178-187, 2021. https://doi.org/10.21533/pen.v10i1.2577 [12] W.-S. Chen and R.-H. Jeng, "A new patch-based LBP with adaptive weights for gender classification of human face," Journal of the Chinese Institute of Engineers, vol. 43, no. 5, pp. 451-457, 2020. https://doi.org/10.1080/02533839.2020.1751724 [13] M. A. Roa'a, I. A. Aljazaery, and S. K. Al_Dulaimi, and Informatics, "Generation of High Dynamic Range for Enhancing the Panorama Environment," Bulletin of Electrical Engineering, vol. 10, no. 1, 2021. https://doi.org/10.11591/eei.v10i1.2362 [14] A. Swaminathan, M. Chaba, D. K. Sharma, and Y. Chaba, "Gender classification using facial embeddings: A novel approach," Procedia Computer Science, vol. 167, pp. 2634-2642, 2020. https://doi.org/10.1016/j.procs.2020.03.342 [15] S. Kumar, S. Singh, and J. Kumar, "Gender classification using machine learning with multi- feature method," in 2019 IEEE 9th Annual Computing and Communication Workshop and Conference (CCWC), 2019: IEEE, pp. 0648-0653. https://doi.org/10.1109/CCWC.2019. 8666601 [16] F. H. C. Tivive and A. Bouzerdoum, "A gender recognition system using shunting inhibitory convolutional neural networks," in The 2006 IEEE International Joint Conference on Neural Network Proceedings, 2006: IEEE, pp. 5336-5341. https://doi.org/10.1109/IJCNN.2006. 247311 [17] S. Baluja and H. A. Rowley, "Boosting sex identification performance," 2007. [18] A. Nejatian and G. Sarbishei, "Implementation real-time gender recognition based on facial features using a hybrid neural network Imperialist Competitive Algorithm," in 2017 Iranian Conference on Electrical Engineering (ICEE), 2017: IEEE, pp. 1584-1589. https://doi.org/ 10.1109/IranianCEE.2017.7985298 [19] Z. Xu, H. R. Wu, X. Yu, K. Horadam, and B. Qiu, "Robust shape-feature-vector-based face recognition system," IEEE Transactions on Instrumentation and Measurement, vol. 60, no. 12, pp. 3781-3791, 2011. https://doi.org/10.1109/TIM.2011.2141270 [20] A. Mian, M. Bennamoun, and R. Owens, "An efficient multimodal 2D-3D hybrid approach to automatic face recognition," IEEE transactions on pattern analysis and machine intelligence, vol. 29, no. 11, pp. 1927-1943, 2007. https://doi.org/10.1109/TPAMI.2007. 1105 [21] F. Juefei-Xu, E. Verma, P. Goel, A. Cherodian, and M. Savvides, "Deepgender: Occlusion and low resolution robust facial gender classification via progressively trained convolutional neural networks with attention," in Proceedings of the IEEE conference on computer vision and pattern recognition workshops, 2016, pp. 68-77. https://doi.org/10.1109/CVPRW.2016. 24 [22] Y. Wang and N. Zhang, "Gender classification based on enhanced PCA-SIFT facial features," in 2009 first international conference on information science and engineering, 2009: IEEE, pp. 1262-1265. https://doi.org/10.1109/ICISE.2009.620 [23] M. Vasileiadis, G. Stavropoulos, and D. Tzovaras, "Facial soft biometrics detection on low power devices," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019, pp. 0-0. https://doi.org/10.1109/CVPRW.2019. 00285 [24] M. Duan, K. Li, C. Yang, and K. Li, "A hybrid deep learning CNN–ELM for age and gender classification," Neurocomputing, vol. 275, pp. 448-461, 2018. https://doi.org/10.1016/ j.neucom.2017.08.062 [25] J. S. Nayak and M. Indiramma, "An approach to enhance age invariant face recognition performance based on gender classification," Journal of King Saud University-Computer iJIM ‒ Vol. 17, No. 08, 2023 131 https://doi.org/10.21533/pen.v10i1.2577 https://doi.org/10.1080/02533839.2020.1751724 https://doi.org/10.11591/eei.v10i1.2362 https://doi.org/10.1016/j.procs.2020.03.342 https://doi.org/10.1109/CCWC.2019.8666601 https://doi.org/10.1109/CCWC.2019.8666601 https://doi.org/10.1109/IJCNN.2006.247311 https://doi.org/10.1109/IJCNN.2006.247311 https://doi.org/10.1109/IranianCEE.2017.7985298 https://doi.org/10.1109/IranianCEE.2017.7985298 https://doi.org/10.1109/TIM.2011.2141270 https://doi.org/10.1109/TPAMI.2007.1105 https://doi.org/10.1109/TPAMI.2007.1105 https://doi.org/10.1109/CVPRW.2016.24 https://doi.org/10.1109/CVPRW.2016.24 https://doi.org/10.1109/ICISE.2009.620 https://doi.org/10.1109/CVPRW.2019.00285 https://doi.org/10.1109/CVPRW.2019.00285 https://doi.org/10.1016/j.neucom.2017.08.062 https://doi.org/10.1016/j.neucom.2017.08.062 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… and Information Sciences, vol. 34, no. 8, pp. 5183-5191, 2022. https://doi.org/10.1016/ j.jksuci.2021.01.005 [26] T. K. Sajja and H. K. Kalluri, "Gender classification based on face images of local binary pattern using support vector machine and back propagation neural networks," Adv. Modell. Anal. B, vol. 62, no. 1, pp. 31-35, 2019. https://doi.org/10.18280/ama_b.620105 [27] O. Agbo-Ajala and S. Viriri, "Deeply learned classifiers for age and gender predictions of unfiltered faces," The Scientific World Journal, vol. 2020, 2020. https://doi.org/10.1155/ 2020/1289408 [28] "Face Recognition Technology (FERET)," https://www.nist.gov/programs-projects/face- recognition-technology-feret. [29] "UTKFace," https://susanqq.github.io/UTKFace/. [30] G. Azzopardi, A. Greco, A. Saggese, and M. Vento, "Fusion of domain-specific and trainable features for gender recognition from face images," IEEE access, vol. 6, pp. 24171- 24183, 2018. https://doi.org/10.1109/ACCESS.2018.2823378 [31] S. Dargan, M. Kumar, and S. Tuteja, "PCA-based gender classification system using hybridization of features and classification techniques," Soft Computing, vol. 25, no. 24, pp. 15281-15295, 2021. https://doi.org/10.1007/s00500-021-06118-0 [32] H. T. Salim, and I. A. Aljazaery, "Encryption of Color Image Based on DNA Strand and Exponential Factor," International Journal of Interactive Mobile Technologies (iJIM), 2021. [33] H. Salim, and M. R. Aziz, "Combination of hiding and encryption for data security," International Journal of Interactive Mobile Technologies, Article vol. 14, no. 9, pp. 34-47, 2020. https://doi.org/10.3991/ijim.v14i09.14173 [34] O. Surinta and T. Khamket, "Gender recognition from facial images using local gradient feature descriptors," in 2019 14th International Joint Symposium on Artificial Intelligence and Natural Language Processing (iSAI-NLP), 2019: IEEE, pp. 1-6. https://doi.org/10.1109/ iSAI-NLP48611.2019.9045689 [35] S. L. Agrwal, A. Jhanwar, K. Goswami, S. K. Gupta, and V. Kant, "Facial gender recognition using Gabor-DCT feature extraction," Journal of Statistics and Management Systems, vol. 22, no. 4, pp. 719-728, 2019. https://doi.org/10.1080/09720510.2019.1609728 [36] H. T. Salim, I. A. Aljazaery, J. S. Qateef, A. H. M. Alaidi, and R. a. M. Al_airaji, "Face Patterns Analysis and Recognition System Based on Quantum Neural Network QNN," International Journal of Interactive Mobile Technologies, vol. 16, no. 8, 2022. https://doi.org/10.3991/ijim.v16i08.30107 [37] L. D. Ningrum, B. S. B. Dewantara, and D. M. Sari, "Human Gender Detection from Facial Image using Global and Local Feature," in 2022 International Conference on Electrical and Information Technology (IEIT), 2022: IEEE, pp. 402-407. https://doi.org/10.1109/ IEIT56384.2022.9967892 [38] R. Hammouche, A. Attia, S. Akhrouf, and Z. Akhtar, "Gabor filter bank with deep autoencoder based face recognition system," Expert Systems with Applications, p. 116743, 2022. https://doi.org/10.1016/j.eswa.2022.116743 [39] G. Trivedi and N. N. Pise, "Gender classification and age estimation using neural networks: a survey," International Journal of Computer Applications, vol. 975, p. 8887, 2020. [40] J. S. Hiremath, S. S. Hiremath, S. Kumar, M. S. Chincholi, S. B. Patil, and M. S. Hiremath, "Gender Detection using Facial Features with Support Vector Machine," in 2022 IEEE 2nd International Conference on Mobile Networks and Wireless Communications (ICMNWC), 2022: IEEE, pp. 1-6. https://doi.org/10.1109/ICMNWC56175.2022.10031856 [41] S. Pai and R. Shettigar, "Gender recognition from face images using sift descriptors and trainable features," in Advances in Artificial Intelligence and Data Engineering: Select 132 http://www.i-jim.org https://doi.org/10.1016/j.jksuci.2021.01.005 https://doi.org/10.1016/j.jksuci.2021.01.005 https://doi.org/10.18280/ama_b.620105 https://doi.org/10.1155/2020/1289408 https://doi.org/10.1155/2020/1289408 https://doi.org/10.1109/ACCESS.2018.2823378 https://doi.org/10.1007/s00500-021-06118-0 https://doi.org/10.3991/ijim.v14i09.14173 https://doi.org/10.1109/iSAI-NLP48611.2019.9045689 https://doi.org/10.1109/iSAI-NLP48611.2019.9045689 https://doi.org/10.1080/09720510.2019.1609728 https://doi.org/10.3991/ijim.v16i08.30107 https://doi.org/10.1109/IEIT56384.2022.9967892 https://doi.org/10.1109/IEIT56384.2022.9967892 https://doi.org/10.1016/j.eswa.2022.116743 https://doi.org/10.1109/ICMNWC56175.2022.10031856 Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… Proceedings of AIDE 2019: Springer, 2020, pp. 1173-1186. https://doi.org/10.1007/978- 981-15-3514-7_87 [42] T. Kurita, "Principal component analysis (PCA)," Computer Vision: A Reference Guide, pp. 1-4, 2019. https://doi.org/10.1007/978-3-030-03243-2_649-1 [43] M. J. Al-Dujaili and M. T. Mezeel, "Novel approach for reinforcement the extraction of ECG signal for twin fetuses based on modified BSS," Wireless Personal Communications, vol. 119, no. 3, pp. 2431-2450, 2021. https://doi.org/10.1007/s11277-021-08337-y [44] B. Ghojogh, S. B. Shouraki, H. Mohammadzade, and E. Iranmehr, "A fusion-based gender recognition method using facial images," in Electrical Engineering (ICEE), Iranian Conference on, 2018: IEEE, pp. 1493-1498. https://doi.org/10.1109/ICEE.2018.8472550 [45] M. J. Al Dujaili, A. Ebrahimi-Moghadam, and A. Fatlawi, "Speech emotion recognition based on SVM and KNN classifications fusion," International Journal of Electrical and Computer Engineering, vol. 11, no. 2, p. 1259, 2021. https://doi.org/10.11591/ijece.v11i2. pp1259-1264 [46] M. Payasi and K. Cecil, "LBP and Iris Features based Human Gender Classification using radial Support Vector Machine," in 2021 Fourth International Conference on Electrical, Computer and Communication Technologies (ICECCT), 2021: IEEE, pp. 1-7. https://doi. org/10.1109/ICECCT52121.2021.9616923 [47] S. Ghosh, "Semantic annotation for gender identification using support vector machine," IJAR, vol. 3, no. 4, pp. 147-160, 2017. 7 Authors Mohammed Jawad Al-Dujaili Al-Khazraji awarded B.Sc. degree in communica- tion engineering from University of Al-Furat Al-Awsat Technical, Technical College of Engineering, Najaf, Iraq in 2008 and M.Sc. degree in communication system engi- neering from Ferdowsi university, Iran, in 2018. Currently, he is a member staff at the Department of Electronic and Communication, Faculty of Engineering, University of Kufa, Iraq. His research interest includes the development of Wireless communications and signal processing as well as image, speech processing and radar, 5G, 6G, IOT. He can be contacted at email: Mohammed.challab@uokufa.edu.iq. Haider TH. Salim ALRikabi is presently Asst. Prof and one of the faculty Electri- cal Engineering Department, College of Engineering, Wasit University in Al Kut, Wasit, Iraq. He received his B.Sc. degree in Electrical Engineering in 2006 from the Al Mustansiriya University in Baghdad, Iraq. His M.Sc. degree in Electrical Engineering focusing on Communications Systems from California state university/Fullerton, USA in 2014. He is author, coauthor, and Editor of some national and international journals and conference papers. His current research interests include Communications systems with the mobile generation, Control systems, intelligent technologies, smart cities, Re- newable Energies, signal processing as well as image, speech processing and the Inter- net of Things (IoT). Al Kut city – Hay ALRabee, Wasit, Iraq. The number of articles in national databases – 10, and the number of articles in international databases – 60. He can be contacted at email: hdhiyab@uowasit.edu.iq. Nisreen Khalil Abed is a lecturer in the Engineering College, at the Wasit Univer- sity, Iraq. His area of research focuses on power systems and their applications. Year iJIM ‒ Vol. 17, No. 08, 2023 133 https://doi.org/10.1007/978-981-15-3514-7_87 https://doi.org/10.1007/978-981-15-3514-7_87 https://doi.org/10.1007/978-3-030-03243-2_649-1 https://doi.org/10.1007/s11277-021-08337-y https://doi.org/10.1109/ICEE.2018.8472550 https://doi.org/10.11591/ijece.v11i2.pp1259-1264 https://doi.org/10.11591/ijece.v11i2.pp1259-1264 mailto:Mohammed.challab@uokufa.edu.iq mailto:hdhiyab@uowasit.edu.iq Paper—Gender Recognition of Human from Face Images Using Multi-Class Support Vector Machine… of Graduation 2006 University of Technology/Electromechanical Engineering Depart- ment – Electrical branch Bachelor degree. M.Sc. Year of Graduation 2020 University of Wasit/Electrical engineering department–general branch Master degree. She can be contacted at email: nabed@uowasit.edu.iq. Ibtihal Razaq Niama ALRubeei received the B.Sc Eng. Degree in Electrical Engi- neering from the University of Technology, Iraq in 2010. She is presently an Engineer in College of Engineering, Electrical Engineering Department, Wasit University in Al Kut, Wasit, Iraq. She received M.Sc in Electrical engineering, Ilam University in 2022. Her current research interests include renewable energies, control system, and Smart Technologies, Image Recognition, and IoT. She can be contacted at email: ibtihalal- rubeei82@gmail.com. Article submitted 2023-02-05. Resubmitted 2023-03-03. Final acceptance 2023-03-05. Final version pub- lished as submitted by the authors. 134 http://www.i-jim.org mailto:nabed@uowasit.edu.iq mailto:ibtihalalrubeei82@gmail.com mailto:ibtihalalrubeei82@gmail.com