Printer Friendly

Face Recognition using Similarity Pattern of Image Directional Edge Response.


Automatic face recognition is an interesting and challenging task and is used in people's everyday life, ranging from security and surveillance systems to smartphones, smart homes and intelligent digital photography [1-3]. In the field of biometrics also, face recognition has seen increased usage compared to some other methods such as fingerprint and iris scan. The major reason is that face recognition does not necessarily require the cooperation of the individual to collect a sample. The image can be taken by any image capturing media even without the individual noticing [4].

The generic steps common to most face recognition systems include cropping and normalizing the face image taken from a camera or other input device followed by feature extraction which generates a representation of the facial image which is then fed into a classifier and the classifier labels the image with one of its database images having the closest resemblance with the input facial image [2]. A generic flow diagram is shown in Fig. 1. The key step is the feature extraction and the challenge is to generate a face representation which is stable in inconsistent environment with different changing factors. Real-world facial images have characteristics such as illumination variation, pose and expression variation, occlusion and noise. Moreover, human faces also have an inherent property of subtle changes in structure over time. An effective feature extraction method should generate a face feature which is robust in changing environment. This feature should be consistent for the same person. It should minimize the within-class variance and maximize the between-class variance of the facial images. Some properties of an effective face feature also include the ease with which it can be extracted from a given image and having low dimension length of the feature space to ensure quick classification time [5].

The facial feature extracted using methods proposed till now can be primarily divided into two categories: Global features and local features. Global features are extracted from the entire image without considering any information about the feature locality. Popular methods of extracting global features include Principal Component Analysis (PCA) [6], Independent Component Analysis (ICA) [7], Fast ICA [8], Linear Discriminant Analysis (LDA) [9] and more recently the 2D-PCA [10]. Image frequency features such as dominant frequency features [11] and polar frequency features [12] have also been proposed. Probabilistic approach to face recognition has also been proposed such as Bayesian Face Recognition [13]. It relies on finding the probabilistic similarities between facial images. Global features lack the important discriminatory information of image locality and thus the performance of these approaches degrades in changing environment, i.e., pose, lighting and expression variation [5].

In recent years, local feature based approaches are gaining popularity in face recognition [14]. Local feature based approaches divides the input image to smaller local regions and extracts discriminatory information from the local regions to generate a face feature. Some local feature extraction method select key facial components such as eyes nose and mouth and encode the texture information of those regions. Some examples include the Elastic Bunch Graph Matching (EGBM) [15], Local Feature Analysis (LFA) [16] and Locally Salient Independent Component Analysis (LS-ICA)[17]. EGBM gained much popularity as it first locates fiducial points, such as eyes, nose and mouth, and applies the Gabor wavelet in those areas to extract the feature of each point, called Gabor jets. Each image is represented as a graph, called bunch graph, where each Gabor jet forms the nodes of the graph. Classification is done by comparing the bunch graph of facial images. Some other facial component localization methods include geometric approaches to face recognition. Geometric methods locate key facial components (eyes, nose, mouth, etc.) and use the geometric relationships such as distance, angle and position on the facial components to generate the face representation [18-19]. However, all these approaches require pin-point selection of the salient facial components, something which is very difficult in changing environment especially in the presence of occlusion. So, these approaches do not perform well in challenging environments. A more generic local facial texture encoding methods have also been proposed by using micro-patterns to represent a facial image. Micro-pattern based method encodes information of each region using an encoding scheme. A facial image is divided into smaller sub regions and a filter is applied on each region to extract the local texture information and to generate the face feature the results from each region are combined. Among the micro-pattern based approaches, Local Binary Pattern (LBP) [20] has found much popularity among the researchers. Initially proposed for texture classification, LBP has also been applied in face recognition [21]. LBP generates an 8-bit binary code for each pixel of the image by comparing the intensity of the corresponding pixel with its neighboring pixels. It proved to be simple yet an effective method for extracting facial features. However, LBP is very sensitive to changes in intensity values and generates inconsistent codes in presence of noise and illumination variation. Although some extensions of LBP were proposed [22-24], none of them could completely overcome the challenge of non-monotonic illumination variation. Tran and Briggs [25] took a step forward and proposed Local Ternary Pattern (LTP). It quantizes the intensity values into three levels based on the intensity of the center pixel and a threshold, T, and generates two 8-bit binary codes for each pixel. Although the threshold negated the effect of small fluctuations in the intensity values due to noise and illumination variation, LTP does not perform well in presence of non-monotonic illumination variation and expression variation.

Realizing the sensitivity of pixel intensities to even slight changes, researchers have also proposed methods which encode the intensity values using a filter bank or mask. Deniz et al. [26] proposed an approach using the orientation of pixel gradients. It selects salient regions of the face and calculates the gradient orientation of those regions. Locating salient regions accurately is a challenge in presence of occlusion and pose variation and the proposed method's performance degraded in such situations. Jabid et al. [27] used directional edge responses of facial image to generate a face representation. For each pixel, the edge response of the neighbor cells in the 3x3 local region is calculate and the top K neighbors having the highest edge responses are assigned a value of 1 and the remaining are assigned a value of 0. The generated 8-bit binary code is assigned to the center pixel of each local region. Since Local Directional Pattern (LDP) depends on large differences in edge response, it fails to generate consistent code in smooth or near-smooth regions.

In this paper we propose a new face descriptor, namely, Similarity Pattern of Image Directional Edge Response (SPIDER). It first generates the directional edge response of a given facial image. For each pixel, a cell is defined as the 3x3 neighborhood around it, a histogram is generated containing 8 bins and each bin is the sum of the edge response of the neighbors in one of the 8 respective directions. A block is defined as a 3x3 neighboring cells around each pixel and a similarity pattern is generated by comparing the histogram of the pixel with the histogram of the center pixel of each cell in the block. Each pixel is assigned an 8-bit binary code based on the similarity values generated. The encoded image is used as the face representation. Finally the feature vector generated from the representation is reduced using a dimensionality reduction technique so that the classification time reduces and the proposed method can be used even in systems having millions of database images to choose from in the classification step. All the experiments have been carried out on the FERET database [28] and experiment results have shown that in changing environment with the presence of noise, illumination variation, pose and expression variation and occlusion, the proposed method performs better than the existing approaches. Low feature dimension enables quick classification time, therefore, reducing the feature vector further by dimension reduction technique results in fast classification with minimal loss in recognition accuracy.


LBP was first proposed as a grey-scale invariant operator for texture classification [20]. The idea became popular among researchers for its simplicity and effective performance and they applied the LBP operator in several computer vision tasks [21] [29-31]. The basic LBP operator assigns 8-bit binary code to each pixel by thresholding the gray-scale value of the center with respect to the 8 neighbors in the 3x3 cell around it. Generally, the binary code is converted to a decimal number, before assigning it to each pixel, using equation (1).

[mathematical expression not reproducible] (1)

Here, [i.sub.c] is the gray-scale value of the center pixel ([x.sub.c], [y.sub.c]) and [i.sub.p] is the gray-scale value of the 8 neighbors. The value of f(x) is defined as shown in equation (2).

[mathematical expression not reproducible] (2)

Figure 2 shows how the LBP operator works. The basic LBP operator was later extended to include P equally spaced neighbors located on the circular neighborhood, having radius R, around each pixel. Bilinear interpolation is used to estimate values of neighbor that do not exactly fall on pixels. The size of the neighborhood is varied by changing the value of R. The resulting LBP code consists of a P-bit binary number.

A variation of the LBP code is called uniform LBP code [20]. Uniform codes contain at most two transitions from 0 to 1 and vice versa when the bits are considered in a circular sequence. For example, the binary representation 00000000 and 1111111 consist of zero transitions and 00110000 consist of two transitions and are considered as uniform patterns. Ahonen et al. [21] used the uniform LBP pattern containing at most two transitions for face recognition and found good recognition accuracy in presence of monotonic illumination variation.

However, the LBP operator is very sensitive to even slight changes in gray-scale values and produces inconsistent code. In Fig. 2, if the value of the center pixel changes to 97 then the LBP operator would produce a different binary code (0111010) even though gray-scale values of the 3x3 neighbor cell remained unchanged. In challenging environment including noise and nonmonotonic illumination variation the LBP operator generates inconsistent code and the recognition accuracy degrades [5, 25]. An effective encoding operator should be robust to such changes and generate the same code.


Encoding the micro-level information of edges, spots and other local features in an image is always a great challenge for researchers. The LBP operator uses intensity changes around each pixel to find the information. Some have used gradient magnitude values to encode these features [23]. Also directional edge responses are used for finding out the texture pattern [27]. It is seen that directional edge response encodes more texture information of an image. The proposed SPIDER descriptor is an eight-bit binary code assigned to each pixel of an image by comparing the relative cumulative edge responses value of a particular cell (consisting 3x3 neighboring pixels around the center pixel) with its eight neighboring cells.

For computing the edge responses around eight directions, Kirsch edge detector, Robinson edge detector, Prewitt edge detector, Sobel edge detector etc. can be used [32]. In the proposed method, the Robison mask is used to compute the edge response value of each pixel in eight different orientations ([M.sub.0] - [M.sub.7]). Fig. 3 shows the Robinson masks.

Applying the Robinson mask, eight edge response values are obtained, [m.sub.0], [m.sub.1], [m.sub.2], ..., m7 each representing the edge significance in its respective direction. Fig. 4 shows the edge response calculation process for a particular pixel. Here presence of corner and edge in the image yields high response values in particular directions and smooth regions produce response values of similar magnitudes in each of the 8 directions. To incorporate further local texture information, each of the eight directional edge response values of all the pixels in a 3x3 neighboring region, centered on each pixel (being referred to as a 'cell'), are summed up. Hence, a histogram is generated, consisting of 8 bins containing the accumulated responses, as shown in Fig. 5. The histogram is assigned to the center pixel of the cell.

After generating the histogram of each pixel, the next step is to measure the similarity between different cells to find out the local texture pattern. A 3x3 neighboring cells is defined as a 'block' and the similarity between the center cell and its neighboring cells is calculated. The Chi-square method is used to calculate the similarity value of the histograms of the cells in a block, using the formula below:

[mathematical expression not reproducible] (3)

Here [H.sup.c] is the histogram of the center cell and [H.sup.i] is the histogram of the neighboring i-th cell. Fig. 6 shows an example similarity values generated for a block.

Each of the eight neighbors represents eight directions. The average of the eight dissimilarity values is considered as the threshold value. The cells having a dissimilarity value greater than the average dissimilarity value is assigned a value of 0 and the rest are assigned 1. The eight neighboring cells of a particular block produce an eight-bit binary code, containing information of smooth and highly textured local regions of a facial image. This binary code is converted to decimal value and assigned it to the central pixel of the central cell.

Formally, a SPIDER code can be formulated by the following formula.

SPIDER ([x.sub.c], [y.sub.c]) [7.summation over (p=0)]/s(i) X [2.sup.p] (4)


[mathematical expression not reproducible] (5)

Here [H.sup.c] is the center cell's local histogram and [H.sup.i] is the neighboring i-th cell's local histogram. [[chi].sup.2.sub.avg] is the average of all eight local neighborhood's dissimilarity value. Fig. 7 shows the process of generating a SPIDER code of a particular pixel.


A. Generating SPIDER face feature vector

There are three steps to represent a face using SPIDER descriptor from a raw image. First, SPIDER operator is applied on the face image to extract SPIDER image. The resulting image contains fine details of the image such as edges, corners and information about the high textured regions but histogram computed over the entire image contains only the occurrence of the micro-patterns without any knowledge about their locations. Second, to extract the local information from the face, SPIDER image is partitioned into NxN sub-regions and histogram is extracted from each of the local region. As SPIDER operator assigns an 8-bit binary code to each pixel which is converted into a decimal value, each local region creates a histogram consisting of 256 bins. Third, all histograms are concatenated into one feature vector to build the global representation of the face. Fig. 8 illustrates the SPIDER face feature vector generation process.

B. Dimensionality reduction

Concatenation of the local histograms makes the length of the SPIDER face feature vector having length of NxNx256, which causes higher time complexity for a large number of samples. For reducing the matching time, a dimensionality reduction technique is used to compress the generated feature vector. After experimenting with some existing tools for reducing dimension of data i.e., Principal Component Analysis (PCA) [6], Linear Discriminant Analysis (LDA) [9] and Discrete Cosine Transform (DCT) [33], DCT is chosen as the dimension reduction technique for compressing the feature vector's size. DCT transforms the feature vector from spatial domain to frequency domain. It is seen that most of the feature vector's total energy lies in a small subset of the total DCT coefficients and this subset is retained while the other DCT coefficients are discarded. The DCT coefficients are chosen by taking the significant features where most of the energy lies. The retained subset is used as the final representation of each facial image. This reduces the size of the feature vector while still containing most of the essential information necessary for effective classification. Fig. 9 shows the process of applying DCT to SPIDER face feature and taking significant co-efficient from them.


For face recognition system, there are a large number of classes but only a few numbers of training samples. So instead of using more sophisticated classifier, simple nearest-neighbor classifier performs better in this kind of pattern recognition problem [1]. Several dissimilarity measurement methods are present i.e., Histogram intersection, Log-likelihood statistics and Chi-square statistics ([[chi].sup.2]) [5]. Among them the Chi-square dissimilarity method is used. Chi-square statistics is shown in Equation 3. The feature vector of the input image is compared with the feature vector of each image in the image gallery using Chi-square dissimilarity measurement. The gallery image which gives the least Chi-Square dissimilarity value is chosen as the class of the input. Recent studies show that different facial features have different degree of significance in face recognition like eyes, face outline, mouth are more important for remembering faces while nose plays a relatively unimportant role [1]. The weighted Chi-square dissimilarity measure is used to incorporate the weight of each block's histogram. Assignment of weights to each significant region is shown in Fig. 10. The following formula shows the weighted Chi-square dissimilarity measure.

[mathematical expression not reproducible] (6)

Here, i refer to the region number, j indicates bin number of that region and [w.sub.i] is the weight of region i. [H.sup.1] and [H.sup.2] are histogram of training and testing image respectively. The measure [[chi].sup.2.sub.w]([H.sup.1],[H.sup.2]) represents dissimilarity between these two histograms.

After reducing the feature vector size using DCT, matching of the training and testing image is done using Euclidean distance [33] formula shown below:

[mathematical expression not reproducible] (7)

Here, [H.sub.1] and [H.sub.2] are histogram of training and testing image respectively. The value of D ([H.sup.1], [H.sup.2]) represents dissimilarity between these two histograms.


The recognition accuracy of the proposed SPIDER pattern is tested in accordance with the Colorado State University Face Identification Evaluation System [34] using Images from the FERET face database [28]. To achieve a fair comparison, we implemented the well-known face recognition approaches i.e., Principal Component Analysis (PCA) [10], Elastic Bunch Graph Match (EBGM) [15], Local Binary Pattern (LBP) [21], Local Ternary Pattern (LTP) [25], Local Directional Pattern (LDP) [27] on the same image set and compared it against the proposed method. The experiments were conducted on the 1,196 subjects of FERET face database and found the proposed method performed significantly better than the current methods even in presence of noise, illumination variation, aging, expression variation and occlusion.

A. Experimental Setup

The FERET database consists of a total number of 14,051 frontal face images containing variations in illumination, facial expression, pose angle, aging effects etc. which represents 1,196 particular individuals. These images are divided into five sets, where one set is used for training and the four other for testing the system. Brief descriptions of these sets are shown in Table I.

The system preprocesses an image by cropping out the non-face area from the image. This cropping is done by an elliptical mask which uses the ground-truth data of particular subject's eye, nose and mouth provided in the database. Finally the images are normalized into 105x105 pixel. Fig. 11 shows sample images, each from one of the five sets of FERET database. In the experiments, each image is divided into NxN sub-blocks. Images of fa set are used as gallery image and the other four sets (fb, fc, dupI, dupII) as probe images.

B. Parameters of SPIDER descriptor

To optimize the performance of SPIDER based face recognition system, some parameters tunings were done, which includes value of Threshold K and number of sub-blocks (NxN). After performing several experiments, the average dissimilarity value of the neighborhood pixel is used as the threshold and dividing the image into 7x7 sub-blocks because they provide the highest recognition accuracy.

Table II shows the accuracy while varying the threshold value. From the table it can be deduced that using the average dissimilarity value in a block as the threshold gives the best performance. In a region where the dissimilarity values vary by a large margin, using a static threshold method, such as giving the lowest 3 dissimilarity value cells a value of 1 would generate the same pattern as in a region where the dissimilarity values are very close. However, the textures in the two regions are different which cannot be encoded in the pattern when using a static threshold method. Averaging gives a dynamic allocation of threshold value where the encoded pattern would differ for a high textured region when compared with the one in a low textured region.

Generating the face representation without dividing the image in sub-blocks only gives the frequency of each code in the image and does not provide any information about the location of each code. Two completely different images having the same code frequency would be labeled as the same image which obviously would be a misclassification. Fig. 12 shows the performance analysis according to different sub-division of the SPIDER image. The accuracy for a single division gives a very low accuracy and once some degree of partitioning is introduced, the accuracy changes drastically because the location of each pattern is also incorporated in the feature representation. Increasing the number of subdivision increases the local information and thus provides better accuracy but on the other hand increasing the number of subdivisions also increases the feature length and the classification time. A trade-off between time and accuracy has to be made and the 5x5 subdivision may be a good choice. However, in these days computers have high-end processor and computational power is not a big barrier. So we have finally decided to select 7x7 sub-division in order to achieve higher accuracy. Furthermore, for maintaining the time complexity we have used data reduction technique to filter out the effective features.

For reducing the feature vector size, DCT is used. But choosing the number of significant coefficient is an important factor. Fig. 13 shows the classification accuracy while varying the number of DCT coefficients that are retained. Keeping too little coefficients would result in loss of important between-class and within-class discriminatory information and would reduce the accuracy of the system. As more DCT coefficients are retained the accuracy improves but from the classification time point-of-view, increasing the DCT coefficients makes the system computationally expensive. The number of DCT coefficients chosen should be one which reduces the computation time and at the same time gives a relatively accurate classification. From, the graph it can be seen that using 512 coefficients satisfies the both conditions.

C. Performance Evaluation

The comparison of SPIDER with other existing method is shown in Table III. SPIDER outperforms LBP on all the probe sets, fb, fc, dupI and dupII because LBP uses the intensity values to generate the bit pattern and intensity values would give varying pattern in presence of noise and illumination variation. The use of directional edge response makes SPIDER more stable in noise and lighting changes. The performance of SPIDER is also comparatively better than LDP. While LDP uses a static threshold value, SPIDER uses the dynamic averaging method for thresholding and thus generates stable codes. This gives better result when compared to the other methods especially in the challenging probe sets of the FERET database namely, fc, dupI and dupII. Also performance of SPIDER is also compared with some other past standard methods and it shows that SPIDER performs better than those methods. PCA is a holistic method where the local information of the image can't be extracted. EGBM heavily relies on the location of key facial components before apply the Gabor wavelet to generate the face representation. However, in changing environment the localizing step cannot be done accurately especially in presence of occlusion and expression variation. In such cases the bunch graph generated is not accurate and the performance degrades. SPIDER does not rely on any component localization and thus generates consistent face representation even in presence of such changing factors.

The feature representaiton based on SPIDER is more robust than other existing methods in presence of noise. Therefore, to investigate the sensivity of SPIDER, LDP and LBP to noise, an experiment was conducted by applying Gaussian white noise with different noise variance ([sigma] = 0, 2, 4, 6, 8, 10) to all images of FERET database except the gallery image set, fa. Fig. 14 illustrates the average recognition rates of the four probe sets against different Gaussian noise levels and it can be observed that SPIDER outperforms the existing methods.

It can be deduced from the performance analysis that SPIDER performs better than the existing methods. But it is also noticable that for large number of dataset, the matching time becomes huge which sometimes degrades the performance of the system. So, DCT is applied to the SPIDER face descriptor which creates feature vector of lesser dimension. The performance using this feature vector is also shown in Table III. It is seen that the accuracy of SPIDER+DCT is less than SPIDER descriptor because information is lost due to feature dimension reduction. For smaller dataset using SPIDER, without reducing the feature vector length, is a better solution but for larger dataset it is optimal to use SPIDER+DCT since it will save a lot of time during matching. The matching time comparison is shown in Fig. 15. It illustrates that use of DCT with SPIDER makes the histogram matching time more faster than other methods.


In this paper we have proposed a new face descriptor, SPIDER, for extracting feature representation from a given facial image. Experiments carried out on FERET database have shown superior performance of SPIDER descriptor compared to the existing face descriptors. The use of edge responses and region similarity pattern makes the method robust to noise, pose, expression and age variation. Applying dimension reduction techniques on the SPIDER feature vector is seen to have no drastic impact on the recognition rate. Currently the proposed method is tested on static frontal facial image and in future we plan to implement our proposed method for three-dimensional (3D) face recognition from video sequences.


[1] M. D. Kelly, "Visual Identification of people by computer", PhD Thesis, Stanford University, Stanford, CA, USA, 1971

[2] W. Zhao, R. Chellappa, P. J. Phillips and A. Rosenfeld, "Face Recognition: A Literature Survey", ACM Computing Surveys, vol. 35, no. 4, pp. 399-458, 2003.

[3] D. S. Kim, I. J. Jeon, S. Y. Lee, P. K. Rhee and D. J. Chung, "Embedded face recognition based on fast genetic algorithm for intelligent digital photography," IEEE Transactions on Consumer Electronics, vol. 52, no. 3, pp.726-734, 2006.

[4] F. Zuo and P. H. N. de With, "Real-time embedded face recognition for smart home", IEEE Transaction on Consumer Electronics, vol. 51, no. 1, pp. 183-190, 2005.

[5] T. Jabid, M. H. Kabir, and O. Chae, "Local Directional Pattern (LDP) for face recognition", International Journal of Innovative Computing, Information and Conrol, vol. 8, no. 4, pp. 2423-2437, 2012

[6] M. A. Turk and A. P. Pentland, "Face recognition using eigenfaces," International Conference on Computer Vision and Pattern Recognition, IEEE, 1991, pp. 586-591.

[7] J. R. Movellan, M.S. Bartlett and T. J. Senjnowski, "Face recognition by independent component analysis," IEEE Transactions on Neural Networks, vol. 13, pp.1450-1465, 2002.

[8] C. Zhou, X. Wei, Q. Zhang and B. Xiao, "Image reconstruction for face recognition based on fast ICA", International Journal of Innovative Computing, Information and Control, vol. 4, no. 7, pp. 1723-1732, 2008.

[9] K. Etemad and R. Chellappa, "Discriminant Analysis for recognition of human face images," Journal of the optical Society of America, vol.14, pp. 1724-1733, 1997.

[10] A. F. Frangi, J. Yang, D. Zhang and J. Y. Yang, "Two-dimensional PCA: a new approach to appearance-based face representation and recognition," IEEE Transaction on Pattern Analysis and Machine Intelligence, vol. 26, pp. 131-137, 2004.

[11] I. G. P. Wijaya, K. Uchimura and Z. Hu, "Face recognition based on dominant frequency features and multi-resolution metric," International Journal of Innovative Computing, Information and Control, vol.5, no.3, pp. 641-651, 2009.

[12] Y. Zana and R. M. C. Jr., "Face recognition based on polar frequency features," ACM Transactions on Applied Perception, vol.3, no.1, pp. 62-82, 2006.

[13] B. Moghaddam, T. Jebara and A. Pentland, "Bayesian Face Recognition," Pattern Recognition, vol. 33, no. 11, pp. 1771-1782, 2000.

[14] J. Zhou, Q. Ji and G. Nagy, "A comparative study of local matching approach for face recognition," IEEE Transactions on Image Processing, vol. 16, no. 10, pp. 2617-2628, 2007.

[15] L. Wiskott, J. M.Fellous, N. Kuiger and C. von der Malsburg, "Face recognition using elastic bunch graph matching," IEEE Transaction on Pattern Analysis and Machine Intelligence, vol. 19, no. 7, pp. 775-779, 1997.

[16] P. S. Penev and J. Atick, "Local feature analysis: A general statistical theory for object representation", Network: Computation in Neural Systems, vol. 7, no. 3, pp. 477-500, 1996.

[17] J. Kim, J. Choi and J. Yi, "Face recognition based on locally salient ICA information", Biometric Authentication Workshop, pp. 1-9, 2004.

[18] V. V. Starovoitov, D. I. Samal and D. V. Biriliuk, "Three approaches for face recognition," International Conference on Pattern Recognition and Image Analysis, 2002, pp. 707-7011.

[19] R. Jafri and H. R. Arabnia, "A survey of face recognition techniques," Journal of Information Processing Systems, vol. 5, no. 2, pp. 41-68, 2009.

[20] T. Ojala, M. Pietikainen and T. Maenpaa, "Multiresolution gray-scale and rotation invariant texture classification with local binary patterns," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 7, pp.971-987, 2002.

[21] T. Ahonen, A. Hadid and M. Pietikainen, "Face description with local binary patterns: Application to face recognition," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.28, no.12, pp.2037-2041, 2006.

[22] R. Mattivi and L. Shao, "Human action recognition using LBP-TOP as sparse spatio-temporal feature descriptor", Proc. of International Conference of Computer Analysis of Image and Pattern, pp. 740-747,

[23] S. Zhao, Y. Gao and B. Zhang, "Sobel-LBP," IEEE International Conference on Image Processing, 2008, pp. 2144-2147.

[24] F. Ahmed, E. Hossain, A. S. M. H. Bari, M. S. Hossen, "Compound Local Binary Pattern (CLBP) for rotation invariant texture classification", International Journal of Computer Applications, vol. 33, no. 6, pp. 5-10, 2011.

[25] X. Tran and B. Triggs, "Enhanced local texture feature sets for face recognition under difficult lighting conditions," Analysis and Modeling of Faces and Gestures, pp. 168-182, 2007.

[26] O. Deniz, G. Bueno, J. Salido and F. de la Torre, "Face recognition using Histogram of Oriented Gradients," Pattern Recognition Letters, vol. 32, no. 12, pp. 1598-1603, 2011.

[27] T. Jabid, M. H. Kabir and O. Chae, "Local Directional Pattern (LDP) for face recognition," International Conference on Consumer Electronics, pp. 329-330, 2010.

[28] J. Huang, P. J. Phillips, H. Wechsler and P. Rauss, "The FERET database and evaluation procedure for face recognition algorithms," Image and Vision Computing, vol.16, no.5, pp. 295-306, 1998.

[29] Y. Mu, S. Yan, Y. Liu, S. T.Huang and B. Zhou, "Discriminative local binary patterns for human detection in personal album," IEEE Conference on Computer Vision and Pattern Recognition, 2008, pp. 1-8.

[30] Z. Sun, T. Tan and X. Qiu, "Graph matching iris image blocks with local binary pattern," Advances in Biometrics, 2005, pp. 366-372.

[31] X. Wang, H. Gong, H. Zhang, B .Li and Z. Zhuang, "Palm print Identification using Boosting Local Binary Pattern," International Conference on Pattern Recognition, 2006, pp. 503-506.

[32] R. C. Gonzalez, Digital Image Processing, Pearson, 2007

[33] Z. P. Rod, R. Adams, and H. Bolouri Dimensionality Reduction of Face Images Using Discrete Cosine Transforms for Recognition, IEEE Conference of Computer Vision and Pattern Recognition,

[34] J. Beveridge, D. Bolme, B. Draper, and M. Teixeira, "The CSU face identification evaluation system: Its purpose, features, and structure", Machine Vision and Applications, vol. 16, no. 2, pp. 128-138, 2005.

Farhan BASHAR, Asif KHAN, Faisal AHMED, Hasanul KABIR

Department of Computer Science and Engineering, Islamic University of Technology, Dhaka, Bangladesh

{shadhon, asifkhan, fahmed, hasanul}

Digital Object Identifier 10.4316/AECE.2014.01011

Set    Property                                    No. of Images

fa     Gallery set used for training the system    1196
fb     Alternative facial expressions of subjects  1195
       than in fa
fc     Photos were taken under different lighting   194
dupI   Photos were taken later in time              722
dupII  Subset of dupI taken at least a year after   234
       the corresponding gallery image


K        fb    fc    dupI  dupII

1        0.88  0.82  0.71  0.68
2        0.92  0.84  0.72  0.70
3        0.96  0.86  0.75  0.73
4        0.94  0.85  0.74  0.71
Average  0.98  0.88  0.76  0.74


Method                fb    fc    dupI  dupII

SPIDER (un-weighted)  0.96  0.85  0.71  0.68
SPIDER (weighted)     0.98  0.88  0.76  0.74
SPIDER + DCT          0.95  0.84  0.73  0.71
LDP                   0.96  0.82  0.72  0.69
LTP                   0.95  0.80  0.70  0.67
LBP                   0.95  0.79  0.66  0.64
PCA                   0.85  0.65  0.45  0.22
EBGM                  0.88  0.42  0.46  0.24
COPYRIGHT 2014 Stefan cel Mare University of Suceava
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2014 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Bashar, Farhan; Khan, Asif; Ahmed, Faisal; Kabir, Hasanul
Publication:Advances in Electrical and Computer Engineering
Date:Feb 1, 2014
Previous Article:Fast Decision Tree Algorithm.
Next Article:Wake-on-a-Schedule: Energy-aware Communication in Wi-Fi Networks.

Terms of use | Privacy policy | Copyright © 2022 Farlex, Inc. | Feedback | For webmasters |