human age-group estimation based on anfis using the hog and lbp features

Upload: anonymous-lsebps

Post on 03-Apr-2018

221 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/29/2019 Human Age-Group Estimation Based On Anfis Using the Hog And Lbp Features

    1/9

    Electrical and Electronics Engineering: An International Journal (ELELIJ) Vol 2, No 1, February 2013

    21

    HUMANAGE-GROUP ESTIMATION BASED ON

    ANFIS USING THE HOG AND LBPFEATURES

    Hamid Moghadam fard1

    , Sohrab Khanmohammadi2

    , Sahraneh Ghaemi3

    andFarshad Samadi

    4

    Control Engineering Department, Faculty of Electrical and Computer Engineering,

    Tabriz University, Tabriz, [email protected]

    [email protected]@tabrizu.ac.ir

    [email protected]

    ABSTRACT

    In this paper, a new age-group estimation method is proposed based on fuzzy inference system. Using this

    method, the facial images could be classified in four groups with more accuracy for classification of

    marginal ages. The proposed algorithm is divided into 3 stages. At first stage, the images are

    preprocessed. Then, the image features are extracted using histograms of oriented gradients (HOG) and

    local binary patterns (LBP). Finally, by using neural networks and adaptive neuro-fuzzy inference system

    (ANFIS) the age-groups are classified. Experimental results show that by using this method, the age-

    group classification is done with better performance and the accuracy of age estimation is improved.

    KEYWORDS

    Age-group Classification, Facial Features Extraction, ANFIS, Neural Networks, Histograms of Oriented

    Gradients, Local Binary Patterns

    1.INTRODUCTION

    Human face conveys important perceptible information related to individual traits [1]. Using

    this information some characteristics of people such as identity, gender, age and ethnic origin

    can be recognized. One of the important features of the human is age. Age estimation can beused for finding the missing persons, analyzing the corps, analyzing the criminal persons, etc.

    Many researches are done in extracting the facial information. Among the researches focused on

    age, age estimation is more complicated. One of the key problems in age estimation is that, it isnot a standard classification problem. It means that, it can be taken as a multiclass classification

    or a regression problem. Also, the age progression displayed on faces is uncontrollable and

    personalized [1], [2], [3], [4].

    In recent years, some researches are performed in age estimation field. Kwan and Lobbo [5]

    classified gray-scale images to three groups by using the anthropometric model.Horng et al. [6]proposed another approach that categorized 230 images into four age groups. They extracted

    two geometric features and three wrinkle features; for classification, they considered two back-propagation neural networks. Another estimation method that is called aging pattern subspace(AGES) is proposed by Geng et al [4]. In their proposed method, the sequence of a persons

    images in different ages was used for aging process modeling.

  • 7/29/2019 Human Age-Group Estimation Based On Anfis Using the Hog And Lbp Features

    2/9

    Electrical and Electronics Engineering: An International Journal (ELELIJ) Vol 2, No 1, February 2013

    22

    Lanitis et al. [7] proposed a method to explore the effects of aging on facial appearances. Theyused active appearance model (AAM) to build aging functions. The local binary pattern (LBP)

    method is used on age classification by Gunay and Nabiev[8]. In their method, the face image isdivided to small regions and spatial LBP histograms were produced for each face. Thesehistograms are used to classify the images on age groups. An age classification method was

    proposed by Gao and Ai [9]; they introduced fuzzy-LDA method to solve the age ambiguity

    problem. The other age estimation method was proposed by Choi et al. [10], combining features

    and hierarchical classifiers. In their proposed method, a region specific Gabar filters and LBFmethod are used to extract the wrinkles and skin features, respectively. Hajizadeh and

    Ebrahiminezhad [11] proposed a method for age-group classification using HOG as facialfeatures. Their experimental results demonstrated that in their proposed method, the recognition

    rate for age classification is 87.025%.

    In this paper a fuzzy inference system is proposed to age-group estimation. Using thisalgorithm, the facial images are classified to four groups: underage, young, middle age and old.

    The classification algorithm is divided to three stages: pre-processing, features extraction andclassification. In the proposed method, the image features extraction is performed using HOG

    and LBP. Neural networks and ANFIS are used for age-group classification. Experimentalresults demonstrate the better performance and accuracy of the proposed method.

    The paper is organized as follows: The proposed algorithm for age group estimation thatincludes 3 stages, pre-processing, feature extraction and classification, is explained in section 2.

    The experimental results and their analysis are illustrated and discussed in section 3. Finally thepaper is concluded in section 4.

    2.PROPOSED ALGORITHM FOR AGE-GROUP ESTIMATION

    The proposed algorithm for age group classification is explained in this section. The algorithm

    contains three stages including preprocessing, features extraction and classification. The totalflowchart of the proposed method is illustrated in Fig. 1. These stages are discussed infollowing sections.

    Figure 1. The total flowchart of the proposed method

    2.1. Pre-processing

    The first step for age group classification is preprocessing. In this stage, the RGB image isconverted to a gray-scale image. By using a neural network, face is detected and the horizontal

  • 7/29/2019 Human Age-Group Estimation Based On Anfis Using the Hog And Lbp Features

    3/9

    Electrical and Electronics Engineering: An International Journal (ELELIJ) Vol 2, No 1, February 2013

    23

    and vertical projections of the gray scale image are calculated to find the eyes position. Theseprojections are also used to find the position of pupils. Eyebrows, eyes, nose and mouth areas

    are indicated as the minimum points in vertical projection. With assumption that the eyes are atthe top half of the image and using vertical projection, the line passing through the eyes isindicated [8]. Then the area is determined as the eye area (Fig. 2).

    Figure 2. Vertical and horizontal projections [8]

    2.2. Features extraction

    In this section, the proposed method for features extraction is explained. In this stage, the LBP

    and HOG methods are used to analyze the skin texture and skin wrinkles, respectively.

    2.2.1. Skin texture analysis

    As people getting older, facial blemishes such as freckles and age spots increase on the skin[10]. Also the skin is affected by ultra violet rays from sun. Some regions of the face such as

    cheeks and the bridge of the nose are affected by sunlight, more than other areas. So these areas

    of the image should be cropped to analyse. To crop these regions, the distance between eyes(Deye) is considered as references. The coordinate of these regions are as follow:Cheeks: A rectangular region with 0.33Deye width and 0.4Deye height. The distance between

    the regions upper side and the eye in the same side is 0.4Deye.

    Bridge of nose: A rectangular region with 0.2Deye width and 0.33Deye height. The distance

    between its center and eyes line is 0.5Deye.

    Figure 3. The cropped areas for skin texture analysis

  • 7/29/2019 Human Age-Group Estimation Based On Anfis Using the Hog And Lbp Features

    4/9

    Electrical and Electronics Engineering: An International Journal (ELELIJ) Vol 2, No 1, February 2013

    24

    The cropped areas are illustrated in Fig. 3. After cropping, skin texture analysis is done usingLBP. LBP operator assigns a code to each pixel by comparing it with its neighboring pixels [8].

    The LBP code is created for each pixel using following equation:

    ( ) ( ) ( )