C3 Effective features inspired from Ventral and dorsal stream of visual cortex for view independent face recognition

Similar documents
Texture Feature Extraction Inspired by Natural Vision System and HMAX Algorithm

A Binarization Algorithm specialized on Document Images and Photos

Content Based Image Retrieval Using 2-D Discrete Wavelet with Texture Feature with Different Classifiers

Edge Detection in Noisy Images Using the Support Vector Machines

A Fast Visual Tracking Algorithm Based on Circle Pixels Matching

Skew Angle Estimation and Correction of Hand Written, Textual and Large areas of Non-Textual Document Images: A Novel Approach

Detection of an Object by using Principal Component Analysis

Learning the Kernel Parameters in Kernel Minimum Distance Classifier

FEATURE EXTRACTION. Dr. K.Vijayarekha. Associate Dean School of Electrical and Electronics Engineering SASTRA University, Thanjavur

An Image Fusion Approach Based on Segmentation Region

Face Recognition Based on SVM and 2DPCA

SLAM Summer School 2006 Practical 2: SLAM using Monocular Vision

Cluster Analysis of Electrical Behavior

EYE CENTER LOCALIZATION ON A FACIAL IMAGE BASED ON MULTI-BLOCK LOCAL BINARY PATTERNS

Local Quaternary Patterns and Feature Local Quaternary Patterns

User Authentication Based On Behavioral Mouse Dynamics Biometrics

Improvement of Spatial Resolution Using BlockMatching Based Motion Estimation and Frame. Integration

Simulation: Solving Dynamic Models ABE 5646 Week 11 Chapter 2, Spring 2010

Problem Definitions and Evaluation Criteria for Computational Expensive Optimization

Outline. Discriminative classifiers for image recognition. Where in the World? A nearest neighbor recognition example 4/14/2011. CS 376 Lecture 22 1

Support Vector Machines

A Fast Content-Based Multimedia Retrieval Technique Using Compressed Data

Lecture 5: Multilayer Perceptrons

Learning-based License Plate Detection on Edge Features

Machine Learning 9. week

Parallelism for Nested Loops with Non-uniform and Flow Dependences

Fast Feature Value Searching for Face Detection

Outline. Self-Organizing Maps (SOM) US Hebbian Learning, Cntd. The learning rule is Hebbian like:

Comparing Image Representations for Training a Convolutional Neural Network to Classify Gender

Classifying Acoustic Transient Signals Using Artificial Intelligence

Gender Classification using Interlaced Derivative Patterns

Face Recognition using 3D Directional Corner Points

Corner-Based Image Alignment using Pyramid Structure with Gradient Vector Similarity

SHAPE RECOGNITION METHOD BASED ON THE k-nearest NEIGHBOR RULE

Palmprint Feature Extraction Using 2-D Gabor Filters

Shape Representation Robust to the Sketching Order Using Distance Map and Direction Histogram

PCA Based Gait Segmentation

Efficient Segmentation and Classification of Remote Sensing Image Using Local Self Similarity

A Modified Median Filter for the Removal of Impulse Noise Based on the Support Vector Machines

MULTISPECTRAL IMAGES CLASSIFICATION BASED ON KLT AND ATR AUTOMATIC TARGET RECOGNITION

Vol. 5, No. 3 March 2014 ISSN Journal of Emerging Trends in Computing and Information Sciences CIS Journal. All rights reserved.

On Modeling Variations For Face Authentication

TN348: Openlab Module - Colocalization

Research and Application of Fingerprint Recognition Based on MATLAB

The Research of Support Vector Machine in Agricultural Data Classification

Classifier Selection Based on Data Complexity Measures *

SRBIR: Semantic Region Based Image Retrieval by Extracting the Dominant Region and Semantic Learning

Learning a Class-Specific Dictionary for Facial Expression Recognition

Classification of Face Images Based on Gender using Dimensionality Reduction Techniques and SVM

A Clustering Algorithm for Key Frame Extraction Based on Density Peak

The Study of Remote Sensing Image Classification Based on Support Vector Machine

A high precision collaborative vision measurement of gear chamfering profile

Lobachevsky State University of Nizhni Novgorod. Polyhedron. Quick Start Guide

Face Detection with Deep Learning

Comparison Study of Textural Descriptors for Training Neural Network Classifiers

Face Recognition University at Buffalo CSE666 Lecture Slides Resources:

A CALCULATION METHOD OF DEEP WEB ENTITIES RECOGNITION

A COMBINED APPROACH USING TEXTURAL AND GEOMETRICAL FEATURES FOR FACE RECOGNITION

Invariant Shape Object Recognition Using B-Spline, Cardinal Spline, and Genetic Algorithm

Classifier Swarms for Human Detection in Infrared Imagery

Using Fuzzy Logic to Enhance the Large Size Remote Sensing Images

Feature Reduction and Selection

A Gradient Difference based Technique for Video Text Detection

An Efficient Face Detection Method Using Adaboost and Facial Parts

Journal of Chemical and Pharmaceutical Research, 2014, 6(6): Research Article. A selective ensemble classification method on microarray data

Hybrid Non-Blind Color Image Watermarking

A Gradient Difference based Technique for Video Text Detection

Computer Aided Drafting, Design and Manufacturing Volume 25, Number 2, June 2015, Page 14

A fast algorithm for color image segmentation

Mining Image Features in an Automatic Two- Dimensional Shape Recognition System

PRÉSENTATIONS DE PROJETS

Enhanced Face Detection Technique Based on Color Correction Approach and SMQT Features

Term Weighting Classification System Using the Chi-square Statistic for the Classification Subtask at NTCIR-6 Patent Retrieval Task

Face Recognition Method Based on Within-class Clustering SVM

12/2/2009. Announcements. Parametric / Non-parametric. Case-Based Reasoning. Nearest-Neighbor on Images. Nearest-Neighbor Classification

Shape-adaptive DCT and Its Application in Region-based Image Coding

COMPLEX WAVELET TRANSFORM-BASED COLOR INDEXING FOR CONTENT-BASED IMAGE RETRIEVAL

Load Balancing for Hex-Cell Interconnection Network

Brushlet Features for Texture Image Retrieval

The Research of Ellipse Parameter Fitting Algorithm of Ultrasonic Imaging Logging in the Casing Hole

Medical X-ray Image Classification Using Gabor-Based CS-Local Binary Patterns

A Bilinear Model for Sparse Coding

ALEXNET FEATURE EXTRACTION AND MULTI-KERNEL LEARNING FOR OBJECT- ORIENTED CLASSIFICATION

A Deflected Grid-based Algorithm for Clustering Analysis

A PATTERN RECOGNITION APPROACH TO IMAGE SEGMENTATION

Research of Image Recognition Algorithm Based on Depth Learning

Human Face Recognition Using Generalized. Kernel Fisher Discriminant

The Comparison of Calibration Method of Binocular Stereo Vision System Ke Zhang a *, Zhao Gao b

Single vs. Population Cell Coding: Gaze Movement Control in Target Search

Image Representation & Visualization Basic Imaging Algorithms Shape Representation and Analysis. outline

A New Feature of Uniformity of Image Texture Directions Coinciding with the Human Eyes Perception 1

Real-time Motion Capture System Using One Video Camera Based on Color and Edge Distribution

Modular PCA Face Recognition Based on Weighted Average

A Novel Adaptive Descriptor Algorithm for Ternary Pattern Textures

Collaboratively Regularized Nearest Points for Set Based Recognition

Histogram of Template for Pedestrian Detection

A Background Subtraction for a Vision-based User Interface *

Network Intrusion Detection Based on PSO-SVM

Improved SIFT-Features Matching for Object Recognition

Scale Selective Extended Local Binary Pattern For Texture Classification

Transcription:

ACSJ Advances n Computer Scence: an nternatonal Journal, Vol. 5, ssue 1, No.19, January 016 SSN : 3-5157 C3 Effectve features nspred from Ventral and dorsal stream of vsual cortex for vew ndependent face recognton Somayeh Saraf Esmal 1, Kevan Maghool and Al Mote Nasrabad 3 1 Department of Bomedcal Engneerng, Scence and Research Branch, slamc Azad Unversty, Tehran, ran. s.saraf@srbau.ac.r Department of Bomedcal Engneerng, Scence and Research Branch, slamc Azad Unversty, Tehran, ran. k_maghool@srbau.ac.r 3 Department of Bomedcal Engneerng, Faculty of Engneerng, Shahed Unversty, Tehran, ran. nasrabad@shahed.ac.r Abstract Ths paper presents a model for vew ndependent recognton usng features bologcally nspred from dorsal and ventral stream of vsual cortex. The presented model s based on the C3 features nspred from Ventral stream and tt's vsual attenton model nspred from the dorsal stream of vsual cortex. The C3 features, whch are based on the hgher layer of the HMAX of the ventral stream of vsual cortex, are modfed to extract mportant features from faces n varous vewponts. By tt's vsual attenton model, vsual attenton ponts are detected from faces n varous vews of faces. Effectve features are extracted from these vsual attenton ponts and the vew ndependent C3 effectve features (C3EFs are created from faces. These C3EFs are used to dstnct mult-classes of dfferent subjects n varous vews of faces. The presented model s tested usng FERET face datasets wth faces n varous vews of faces, and compared wth C features (CSMFs and C3 features of standard HMAX model (C3SMFs. The results llustrated that our presented vew ndependent face recognton model has hgh accuracy and speed n comparson wth standard model features, and can recognze faces n varous vews by 97% accuracy. Keywords: vew ndependent face recognton, HMAX model, tt's vsual attenton model, C3 effectve features, ventral stream, dorsal stream. 1. ntroducton n recent years, provdng computatonal algorthms for face recognton n these dfferent stuatons especally n varous vews of faces have been the fundamental ssues [1-3]. Computatonal vew ndependent face recognton s one challengng work that human vsual system can do t wth hgh-speed performance, easly. Thus based on need, recently the study of bran mechansms of the human vsual system have been more consdered. The vsual system s organzed n two functonally specalzed processng pathways n the vsual cortex. One pathway (from the prmary vsual cortex to the paretal cortex for controllng eye movements and vsual attenton s named dorsal stream, and other pathway (from the prmary vsual cortex towards the nferor temporal lobe ncludng V 1, V, V 4, Posteror nfer temporal (PT, Anteror nfer temporal (AT and Fusform Face Area (FFA s named ventral stream, whch processes detal of objects and faces n dfferent condtons [4-6]. The dorsal and ventral streams are not completely ndependent and there are nteractons. For example, area V 4 s nterconnected wth some vsual attenton areas n dorsal stream [7-9]. Partal analoges of the ventral stream cortex have been used n many computng models n canoncal computer vson. Among these models, HMAX s one powerful computatonal model that models the object recognton mechansm of human ventral vsual stream n vsual cortex [10-13], that frst proposed by Poggo et al., s based on expermental results n neurobology. The extracted bonspred C features of ths model can recognze and classfy objects based on the mechansm of human ventral vsual stream n vsual cortex. Then Serre et al. establshed HMAX model and consdered learnng ablty of the model for real-world object recognton n [13]. The features extracted by ths model are C standard model features (SMFs. The C SMFs of the ventral stream have been used for mult-class face recognton [14-16]. n recent years, dfferent development models of the ventral stream HMAX model have been presented to enhance the effcency of the model and n all these models, some feature extracton methods are consdered [17-0]. Lebo et al. n [0], extended HMAX model and added new S 3 and C 3 layers. They found that the performance of the model on vew ndependent wthn-category dentfcaton tasks on dfferent objects was ncreased and the C 3 features of the extended HMAX model performed sgnfcantly 1 Copyrght (c 016 Advances n Computer Scence: an nternatonal Journal. All Rghts Reserved.

ACSJ Advances n Computer Scence: an nternatonal Journal, Vol. 5, ssue 1, No.19, January 016 SSN : 3-5157 better than C and t was ndependent to vewponts. Also, there s some vsual attenton model whch based on vsual attenton regons n the dorsal stream of the vsual cortex and these models were appled n many applcatons such as target detecton, object recognton, object segmentaton and robotc localzaton [1-4]. These vsual attenton models use low-level vsual features such as colour, ntensty and orentaton to form salency maps and fnd focus of attenton locatons. The basc computatonal model of vsual attenton was proposed by tt n 1998 whch are the basc model for the most of the new models and used bottom-up vsual cortex features n the where path []. n ths paper, a model based on new C 3EFs nspred from Ventral and dorsal stream of vsual cortex for the goal of vew ndependent face recognton are presented. For expermental analyss, the FERET faces datasets n varous vews of faces are utlzed and vew ndependent face recognton task by the SVM classfers on the new C 3EFs of them are done and they are compared wth the results of other extracted features. The rest of the paper s organzed as follows. Secton llustrates the C SMFs and C 3SMFs features of ventral stream model and, also the vsual attenton model of dorsal stream for detectng mportant facal regons. Our proposed vew ndependent face recognton model s presented n secton 3. Secton 4 s presented expermental evaluaton ncludng dataset descrpton, results and detaled dscussons. n the fnal secton, we sum up wth a concluson.. Materal and methods.1 The C SMFs features of ventral stream model The C SMFs features of HMAX model are nspred by ventral stream of vsual cortex, and t created by four layers (S 1, C 1, S, C [1, 13]. S 1 features resemble the smple cells found n the V 1 area of the prmate vsual cortex and conssts of Gabor flters. C 1 features mtate the complex cells n V 1&V area of cortex and have the same number of feature types (orentatons as S 1. These features pools nearby S 1 features (of the same orentaton to reach the poston and scale nvarance over larger local regons, and as a result can also subsample S 1 to reduce the number of features. The values of C 1 features are the value of the maxmum S 1 features (of that orentaton that comes wthn a max flter [13]. S 1 features mtate the vsual area V 4 and posteror nfer temporal (PT cortex. They contan RBF-lke unts whch tuned to object-parts and compute a functon of the dstance between the nput C 1 patches and the stored prototypes. n human vsual system, these patches correspond to learnng patterns of prevously seen vsual mages and store n the synaptc weghts of the neural cells. The S features learn from the tranng set of K patches ( P 1,..., K wth varous n n szes (n n = 4 4, 8 8, 1 1 and 16 16 and all four orentatons at random postons (Thus a patch P of sze n n contans n n 4 elements. Then S features, actng as Gaussan RBF-unts, compute the smlarty scores (.e., Eucldean dstance between an nput pattern X and the stored prototype P : f ( X exp( X P, wth σ chosen proportonal to patch sze. The C features mtate the nferotemporal cortex (T and perform a max operaton over the whole vsual feld and provde the ntermedate encodng of the stmulus. Thus, for each face mage, the C features vector s computed and used for face recognton. Ths vector has robustness propertes. The lengths of C features vector are equal to the number of random patches extracted from the mages and have the property of shft and scale ndependent.. The C 3SMFs features of ventral stream model n the mplementaton of the S 3 and C 3 layers from developed HMAX model whch was proposed by Lebo et al. n [0], the response of a C cell (assocatng templates w at each poston t was gven by Eq. (1: S 1 exp( n 3 j 1 ( w t x, j Then, the S 3 features correspondng to all layers were extracted and the maxmum of these values were utlzed as (Eq. (. The C 3 features mtate the vew ndependent propertes n the FFA of the T [0]. These orgnal features are named C 3SMFs n ths paper. C 3 max( S 3 n ths paper, we used the C SMFs and C 3SMFs features vector for vew ndependent face recognton and present a model whch can be extracted effectvely C and C 3 features from face mage n dfferent vewponts..3 Vsual attenton model n dorsal stream of vsual cortex n face mages, some facal regons are more attentve and helpful regons to face recognton, such as eyes, nose and mouth that have been demonstrated by the results of psychophyscal studes n paper [5]. Human can fnd dstnctve nformaton from face mages n a short tme and wth hgh accuracy. These detected features have so much local nformaton to recognze the smlarty of face mages and can track and match to the smlar face mages. So, vsual attenton models nspred from human vsual j (1 ( Copyrght (c 016 Advances n Computer Scence: an nternatonal Journal. All Rghts Reserved.

ACSJ Advances n Computer Scence: an nternatonal Journal, Vol. 5, ssue 1, No.19, January 016 SSN : 3-5157 systems n vsual attenton regons of dorsal stream can detect salent ponts from face mages. Vsual attenton tt's model bologcally models the vsual attenton regons n the posteror cortex and specfes the locatons of salent ponts from a colour mage smulatng saccadc eye movements of human vson []. We utlzed ths model n our proposed model to fnd automatcally mportant regons of the face by adjustng ts parameters. 3. Proposed vew ndependent face recognton model n the ventral stream HMAX model, the S features learned from the randomly extracted patches. So, maybe some of the extracted specal features from cropped face mages such as extracted features from the forehead and cheek regons are not useful features. Snce, these features caused CPU usage n the system and make the system very slow achevng the effectve features vector. Then, t seems necessary to detect best features. Also, HMAX model only models the ventral stream and the connectons between the vsual attenton regons n the posteror cortex to the ventral stream are not consdered. n the proposed model, n order to model the human-lke face recognton system, we extract the C and C 3 features from vsual attenton ponts and acheve the C EFs and C 3EFs for vew ndependent face recognton. Generally, by combnng the herarchcal ventral stream features wth vsual attenton model, the feature extracton model for face recognton system s proposed as follows (Fg. 1 llustrates dfferent vsual cortex layers n two dorsal and ventral streams, and also the whole structure of our proposed feature extracton model for vew ndependent face recognton nspred from them: 1 For each face mage, the colour features, ntensty features and orentaton features are extracted (nspred from the regons n the prmary vsual cortex, whch showed by red dashed lne n Fg. 1 as represented n tt's Vsual attenton model []. g b R r r b G r r g B b g b r g Y b R G B 3 (3 colour salency map, ntensty salency map and orentaton salency map are founded as Eq. (4-(9. M f (s represents of F feature map n s scale. F F F ncluded ntensty features, C colour features. A functon of N(0 s used for created normalzaton map where the symbol represents nterpolaton of the coarser mage to the fner scale and pont by pont subtracton. n ths system, c {,3} and s c d,, c, s C, c, s where d {,3}. N( M ( c ( s N( M N( M C ( c ( c C ( s ( c M ( s ( s RG RG BY BY (4 and C are the salency maps of ntensty and colour, respectvely (n Eq. (5. c, s 3 3 c sc 3 3 N( M ( c ( s Cc, s N( M RG ( c BY ( s c sc (5 Also, the Gabor flters whch generated n S 1 are used as orentaton features where O ( c, s, denotes the orentaton salency map of by c and s operaton of scales (Eq. (6. O( c, s, O( c, O( s, 3 3 {0,45,90,135 } c sc N( O( c, s, Then the features are combned by Eq. (7 to create salent ponts (SP [, 4]. SP ( C O /3 (7 Then a wnner take all network (WTN s used to detect N salent ponts and N attenton ponts (nspred from the regons n the dorsal stream of the vsual cortex, whch showed by blue dotted lne n Fg. 1. 3 Create S 1 and C 1 features from each face mage (nspred from the regons n the prmary vsual cortex, whch showed by red dashed lne n Fg. 1. 4 Extract N patches p ( 1,..., N n four orentatons and n n best patch szes from C 1 features of each face mage by usng detected attenton ponts as the central pxel of them to create effectve S features. Durng recognton, from each test face mage, N patches are created as X ( 1,..., N patches and the dstance between the patches p and X are calculated accordng to the Eq. (8. X P V exp( k 1,,..., N (8 k Whch s proportonal to the patch sze and N dmensonal vectors create for each face mage. The set of V ( k 1,,..., N forms S EFs (nspred from the V4 and k (6 3 Copyrght (c 016 Advances n Computer Scence: an nternatonal Journal. All Rghts Reserved.

ACSJ Advances n Computer Scence: an nternatonal Journal, Vol. 5, ssue 1, No.19, January 016 SSN : 3-5157 Fg 1 The structure of proposed vew ndependent face recognton model. PT area n the ventral stream of the vsual cortex, whch showed by dark blue dotted lne, also the nteracton between the V 4 area n ventral stream by the dorsal stream of the vsual cortex has been showed by red arrows n Fg. 1. 5 Obtan C EFs by general maxmum on S EFs to create N -dmensonal C EFs vector from dstnctve regons of faces (nspred from AT area n ventral stream vsual cortex whch showed by dark blue dotted dot n Fg. 1. 6 Create S 3EFs Features on C EFs (the response of a C EFs cell by usng the Eq. (9 to extract S 3EFs correspondng to all layers (nspred from the FFA area n the ventral stream vsual cortex whch showed by dark blue dotted dot n Fg. 1. Durng recognton, from each test face mage, C EFs of them are created as x ( 1,..., N responses and the dstance between these responses and assocatng templates w at each poston t are calculated accordng to the Eq. (9. n 1 S EFs exp( ( 3 w t x, j j (9 j 1 7 Obtan C 3 EFs by usng the maxmum on the S 3EFs as followng (Eq. (10 (nspred from the vew ndependent regons n the FFA area of the ventral stream vsual cortex whch showed by dark blue dotted dot n Fg. 1. C 3 EFs max( S 3 EFs (10 8 Do step 1 to 6 for all mages to extract C 3EFs vector from dstnctve regons of faces. 9 Feed C 3EFs and C EFs vectors to SVM and classfy face mages wth the goal of vew ndependent face recognton. 4 Copyrght (c 016 Advances n Computer Scence: an nternatonal Journal. All Rghts Reserved.

ACSJ Advances n Computer Scence: an nternatonal Journal, Vol. 5, ssue 1, No.19, January 016 SSN : 3-5157 4. Expermental Analyss 4.1 mage Dataset To demonstrate the feasblty of our proposed face recognton system, experments on the subset of the colour FERET database are organzed [6]. The subset contans ten classes (unque subjects of face mages, wth varatons n pose, expressons and scales. t consst 00 face mages (10 classes, each wth 0 mages. The proposed model s tested usng a 10-fold cross valdaton strategy. So that, for each fold 18 mages of each ndvdual (180 face mages are selected as tranng samples and the rest mages of each ndvdual as test mages (0 face mages. n the pre-processng method, all mages are cropped manually to remove complex background and then they are reszed to the dmensons of 140 140. Fg. shows a seres of 10 dfferent classes of people from FERET database whch used n ths work. All experments are performed entrely n a 3 Matlab01 expermental envronment (characterstc of a computer system s ntel core duo processor (.66 GHz and 4 GB RAM. Fg. 3 shows twenty samples of mages from one subject n vared vew ponts and expresson. 4. Results and dscusson n proposng a vew ndependent face recognton model, to acheve the effectve features vector, at frst the salency maps of colour, ntensty and orentatons are extracted and the feature map's weghts of them are adjusted to create the salency maps from face mages and select the attended locatons of salent ponts as mportant regons of them. Fg. 4, shows the orgnal mages of faces wth seventy attenton ponts and salency maps on them. Fg. 4(a-c and also Fg. 4(d-f shows the orgnal face mages of two ndvduals n three stuatons of vewponts. n the salency toolbox, local max s selected for normalzaton. As shown n Fg. 4, by vsual attenton model, mportant and mportant regons of faces such as nose, eyes, lp and mole are selected as attenton ponts however the faces are n vared vewponts. Fg. mages from ten classes of cropped FERET dataset. Fg. 3 mages of one subject n vared vewponts and expressons. Fg. 4 (a To (f are orgnal mages. (g To (l are face mages wth selected sxty attenton ponts. (m To (k are salency maps of face mages. After fndng attenton ponts from face mages, t s necessary to convert colour mages to grey scale for extractng C and C 3 features. The bologcal S 1 features of face mages are created by convolvng wth 64 Gabor flters. So, there are 64 S 1 features for each face mage. Fg. 5 shows these S 1 features for one orgnal grey-scale face mage after convolvng wth 64 Gabor flter. Also, Fg. 6 shows the C 1 features n band 1 and n four orentatons ( 0, 45, 90 and 135. For each orentaton n band 1, there are two 7 7 and 9 9 flters. At frst, Maxmum responses to them are calculated for each pxel wth 8 8 grd szes. Then, the C 1 features are created by maxmum on correspondng pxels from two mages n each orentaton. The C EFs and C 3EFs of the mages are extracted usng the proposed vew ndependent face recognton model that presented n last secton. Then, SVM classfers wth RBF kernel are traned usng these C features vectors and the class labels, mplemented usng LBSVM [7]. n ths approach, an SVM s constructed for each class by 5 Copyrght (c 016 Advances n Computer Scence: an nternatonal Journal. All Rghts Reserved.

ACSJ Advances n Computer Scence: an nternatonal Journal, Vol. 5, ssue 1, No.19, January 016 SSN : 3-5157 dscrmnatng that class aganst the remanng 9 classes. The number of SVMs used n ths approach s 10. n the testng phase, the features are extracted usng proposed method and the classfcaton s done on test data usng the test SVM classfers. specfc of C 3EFs n 1 1 patch sze n comparson wth others patch szes. As n other patch szes (Fgure 8(a, Fgure 8(b and Fgure 8(d, some pxels of these dagonals are not observable on the same background. n Fgure 8 each pxel n vertcal and horzontal of RDMs represents one subject n one vewpont. Fg. 5 S 1 features created by 64 Gabor flters. Fg. 7 Scheme to extract herarchcal features and create RDMs. Fg. 6 C 1 features created n band 1 and n four orentatons. n order to fnd proper szes of patches for extractng features, the representaton of dssmlarty matrces (RDMs are created. For obtanng ths goal, we extract proper prototype patches wth dfferent patch szes (4 4, 8 8, 1 1 and 16 16 and attan the best bologcal features. These features create the RDMs whch have vew ndependent dentty specfc between features of eght vewponts of ten subjects. Scheme to extract herarchcal features from 10 subjects n 8 vewponts and create RDMs s shown n Fg. 7, and Fg. 8 shows a comparson of the created RDMs n dfferent patch szes of C 3EFs n equal extracted features (N=40. The best RDMs can be created wth the extracton prototype patches n 1 1 patch sze from each of the cropped face mages (Fg. 8(c. t shows created RDMs on C 3EFs features n 1 1 patch sze have hgh correlaton n 15 dagonals parallel to the man dagonal that represent the vew-ndependent Fg. 8 RDMs on C 3EFs of 80 face mages (8 vewponts of 10 subjects n (a 4 4 patch sze (b 8 8 patch sze (c 1 1 patch sze (d 16 16 patch sze. 6 Copyrght (c 016 Advances n Computer Scence: an nternatonal Journal. All Rghts Reserved.

ACSJ Advances n Computer Scence: an nternatonal Journal, Vol. 5, ssue 1, No.19, January 016 SSN : 3-5157 Also, n order to demonstrate feasblty of the proposed vew ndependent face recognton model, we compared the performance of the C EFs vector and C 3EFs vector (wth selected attenton ponts of the proposed model n best patch sze wth the performances of the C SMFs and C 3SMFs (wthout attenton ponts. So, N patches of C SMFs, C 3SMFs, C EFs and C 3EFs n best patch sze from the same face of test and tran mages are extracted and fed nto SVM classfers to compare the recognton rate of them by ten-fold cross valdaton. Fg. 9 shows the accuracy rate of face recognton usng SVM classfers on the C SMFs, C EFs and the C 3EFs of proposed model n varous numbers of extractng features. Gven the results n Table 1 and Fg. 9, the recognton rates of C 3EFs and C EFs are better than C SMFs and C SMFs. Also, the results n table 1 show that 80 number (N=80 of the C 3EFs are enough to get good performance (around 97% aganst; 300 number of the C 3SMFs are requred to get ths good face recognton rate. So, face recognton by C SMFs and C 3SMFs needs more extracted features and more extractng tme. The extractng tme s the average computng tme of each face mage for extractng features that obtaned by tc-toc functon n Matlab01 software. For example, the extractng tme of C 3EFs s total computng tme for selectng attenton ponts and extractng C 3 features from them. Gven by the results n the table 1, the extractng tme to enhance the good recognton accuracy near 97% by usng C 3SMFs (300 number of features are more than others. So, by usng C EFs and C 3EFs, the approprate vew ndependent face recognton rate n lesser tme s acheved. From these results, the C EFs of the proposed model showed a margnal mprovement over the C SMFs on FERET face database whch manly deals wth varances n vewponts. The advantages of C features ntolerance to varatons and the advantages of vsual attenton model to select the attenton ponts are complementary to each other and mutually enhancng the C EFs to vew ndependent face recognton. Table 1: Accuracy rates of face recognton usng SVM classfers on extracted dfferent features. Features Recognton accuracy Extractng (mean ± standard tme (Sec devaton% CEFs (N=80 94 ±.108 5.863 C3EFs (N=80 97±.85 6.05 CSMFs (N=80 89±3.944 4.59 C3SMFs (N=80 91±3.16 4.93 CSMFs (N=300 93.5 ±.838 9.64 C3SMFs (N=300 97 ± 3.16 10.357 Recognton Rate (% 100 95 90 85 80 75 70 65 60 55 50 10 0 30 40 60 80 Number of Features (Number of Extracted Patches Fg. 9 The comparsons plot of computed recognton accuracy rate of C 3EFs, C EFs, C SMFs and C SMFs n dfferent number of extracted features. 5. Conclusons CSMFs C3SMFs CEFs C3EFs n ths paper, we descrbed a bologcally-motvated framework for vew ndependent face recognton, whch the proposed C 3 EFs was nspred from the ventral and dorsal stream of cortex. n fact, we proposed a model to extract new vew ndependent features, usng vsual attenton model and ventral stream model for the goal of vew ndependent face recognton. By vsual attenton model, we specfed the set of attenton ponts from salent ponts on a gallery of face mages n vared vewponts and by ventral stream features, we extracted proper vew ndependent features from the set of attenton ponts and then ran SVM classfers on the vectors of features obtaned from the nput mages. Through the expermental results, we proved that the C 3EFs by usng attenton ponts n comparson to the same number of C SMFs mproved the face classfcaton accuracy under varyng facal expressons and vewponts. Lkewse, n the proposed model, face recognton needs lesser tme snce we do not only use feature selecton method on C features, but also upgrade recognton rate wth the approprate number of attenton ponts whch selected from mportant regons of the face. n the future work, we wll propose the model that can detect and recognze mult-classes of faces wth complcated and vared backgrounds. Also n ths study, we used only SVM classfer and dd not examne other classfers. Then, n future we research drectons for classfcaton changes, the use of artfcal neural network and fuzzy classfcaton to enhance the effcency measure. 7 Copyrght (c 016 Advances n Computer Scence: an nternatonal Journal. All Rghts Reserved.

ACSJ Advances n Computer Scence: an nternatonal Journal, Vol. 5, ssue 1, No.19, January 016 SSN : 3-5157 Acknowledgments Support of Department of Bomedcal Engneerng, Scence and Research Branch, slamc Azad Unversty, Tehran, ran s gratefully acknowledged. Ths study was extracted from Ph.D. thess that was done n Department of Bomedcal Engneerng, Scence and Research Branch, slamc Azad Unversty, Tehran, ran. References [1] Z. Fan, and B. Lu, "Mult-Vew Face Recognton wth Mn-Max Modular Support Vector Machnes", n Proceedngs of the 1st nternatonal Conference on Advances n Natural Computaton, Changsha, Chna, 007, pp. 396-399. [] T.K. Km, and J. Kttler, "Desgn and Fuson of Pose- nvarant Face-dentfcaton Experts", EEE Trans. Crcuts and Systems for Vdeo Technology, Vol. 16, No.9, 006, pp. 1096 1106. [3] K. R Sngh., M. A. Zaver, and M. M. Raghuwansh M. M., "llumnaton and Pose nvarant Face Recognton: A techncal Revew", JCSM, Vol., 010, pp. 09 038. [4] R.H. Wurtz, and E.R. Kandel, Central vsual pathways. n: E.R.Kandel, J.H. Schwartz, T.M. Jessell, edtors. Prncples of neural scence. 4th ed. New York: McGraw- Hll, p. 53 547, 000. [5] E. Kobatake, and K. Tanaka, "Neuronal selectvtes to complex object features n the ventral vsual pathway of the macaque cerebral cortex", Journal of Neurophysology, Vol.71, No.3, 1994, pp. 856 867. [6] Ungerleder L. G., J. V. Haxby, " What and where n the human bran", Current Opnon n Neurobology, Vol.4, No., pp. 157-165, 1994. [7] L.G. Ungerleder, and L. Pessoa, "What and where pathways", Scholarpeda, Vol. 3, No.11, 008, pp. 534. [8] J.M. Brown, "Vsual streams and shftng attenton", Progress n Bran Research, Vol. 176, 009, pp. 47-63. [9] R. Farvar, "Dorsal ventral ntegraton n object recognton", Bran Research Revews, Vol.61, No.,, 009, pp. 144 153. [10] M. Resenhuber, and T. Poggo, "Herarchcal Models of Object Recognton n Cortex", nature neuroscence, Vol., No.11, 1999, pp. 1019 105. [11] M. Resenhuber, and T. Poggo, "Models of object recognton," nature neuroscence, Vol. 3, No. Suppl, 000, pp. 1199 104. [1] T. Serre, and M. Kouh, C. Cadeu, U. Knoblch, G. Kreman, and T. Poggo, A Theory of Object Recognton: Computatons and Crcuts n the Feedforward Path of the Ventral Stream n Prmate Vsual Cortex, Techncal Report, MT, Massa- chusetts, USA., 005. [13] T. Serre, and L. Wolf, S. Blesch, S. Blesch, M. Resenhuber, "Robust object recognton wth cortex-lke mechansms", EEE Transactons on Pattern Analyss and Machne ntellgence, Vol.9, No. 3, 007, pp. 411 46. [14] J. La, and W.X. Wang, "Face recognton usng cortex mechansm and svm", n 1st nternatonal conference ntellgent robotcs and applcatons, Wuhan, 008, pp. 65 63. [15] E. Meyers, and L. Wolf, "Usng Bologcally nspred Features for Face Processng", nternatonal Journal of Computer Vson, Vol. 76, 008,pp. 93 104. [16] P. Pramod Kumar, and P. Vadakkepat, "Hand posture and face recognton usng a fuzzy-rough approach", nternatonal Journal of Humanod Robotcs, Vol. 7, No. 3, 010, pp. 331 356. [17] S. Dura-Bernal, T. Wennekers, and S. L. Denham, "Top-Down Feedback n an HMAX-Lke Cortcal Model of Object Percepton Based on Herarchcal Bayesan Networks and Belef Propagaton", PLoS ONE, Vol. 7, No. 11, 01. [18] J. Mutch, and D.G. Lowe, "Object Class Recognton and Localzaton Usng Sparse Features wth Lmted Receptve Felds", nternatonal Journal of Computer Vson, Vol. 80, No. 1, 008, pp. 45-57. [19] C. Thérault, N. Thome, M. Cord, "Extended Codng and Poolng n the HMAX Model", EEE Transactons on mage Processng, Vol., No., 013, pp. 764 777. [0] J. Z. Lebo, J. Mutch, and T. Poggo, "Why the Bran Separates Face Recognton from Object Recognton", n n Advances n Neural nformaton Processng Systems (NPS, Granada, Span, 011. [1] J. Han, and K.N. Ngan, "Unsupervsed extracton of vsual attenton objects n color mages", EEE Transactons on Crcuts and Systems for Vdeo Technology, Vol. 16, No. 1, 006, pp. 141-145. [] L. tt, C. Koch, and E. Nebur, "A model of salencybased vsual-attenton for rapd scene analyss", EEE Transactons on Pattern Analyss and Machne ntellgence, Vol. 0, No. 11, 1998, pp. 154 159. [3] L. tt, and C. Koch, "Computatonal modelng of vsual attenton", Nature Revews Neuroscence, Vol., No. 3, 001, pp. 194-03. [4] D. Walther, and C. Koch, "Modelng attenton to salent proto-objects", Neural Networks, Vol. 19, No. 9, 006, pp. 1395 1407. [5] F. Kano, and M. Tomonaga, "Face scannng n chmpanzees and humans: Contnuty and dscontnuty", Anmal Behavour, Vol. 79, 010, pp. 7. [6] P.J. Phllps, H. Wechsler, J. Huang, P. Rauss, "The FERET database and evaluaton procedure for face recognton algorthms", mage and Vson Computng, Vol. 16, No. 5, 1998, pp. 95-306. [7] C.C. Chang, C.J. Ln, LBSVM: A lbrary for support vector machnes, Avalable at http://www.cse.ntu.edu.tw/cjln/lbsvm. Somayeh Saraf Esmal, was born n 1984. She receved her B.S, M.S. and Ph.D degree n Boelectrc Engneerng from Scence and Research Branch of slamc Azad Unversty, Tehran, ran n 006, 009 and 015 respectvely. Snce 011, she has been a lecturer n the Garmsar Branch of slamc Azad Unversty, ran. Her current man research ncludes Bonspred Computng and Applcatons n face recognton. 8 Copyrght (c 016 Advances n Computer Scence: an nternatonal Journal. All Rghts Reserved.

ACSJ Advances n Computer Scence: an nternatonal Journal, Vol. 5, ssue 1, No.19, January 016 SSN : 3-5157 Kevan Maghool, s the Assstant Professor at Department of Bomedcal Engneerng, Scence and Research Branch, slamc Azad Unversty, Tehran, ran. He was a Ph.D. Scholar at mentoned Department. He receved M.S. degree n Bomedcal Engneerng from Tarbat Modares Unversty, Tehran, ran. He has more than 10 years of experence, ncludng teachng graduate and undergraduate classes. He also leads and teaches modules at B.Sc., M.Sc. and Ph.D. levels n Bomedcal Engneerng. Hs current research nterests are pattern recognton, bometrc and data mnng. Al Mote Nasrabad, receved a B.S. degree n Electronc Engneerng n 1994 and hs M.S. and Ph.D. degrees n Bomedcal Engneerng n 1999 and 004, respectvely, from Amrkabr Unversty of Technology, Tehran, ran. Snce 005, he has been Assocate Professor n the Bomedcal Engneerng Department at Shahed Unversty, n Tehran, ran. Hs current research nterests are n the felds of bomedcal sgnal processng, nonlnear tme seres analyss and evolutonary algorthms. Partcular applcatons nclude: EEG sgnal processng n mental task actvtes, hypnoss, BC, epleptc sezure predcton and vsual attenton models. 9 Copyrght (c 016 Advances n Computer Scence: an nternatonal Journal. All Rghts Reserved.