Face Recognition using Hough Peaks extracted from the significant blocks of the Gradient Image

Transcription

1 Face Recognition using Hough Peaks extracted rom the signiicant blocks o the Gradient Image Arindam Kar 1, Debotosh Bhattacharjee, Dipak Kumar Basu, Mita Nasipuri, Mahantapas Kundu 1 Indian Statistical Institute, Kolkata , India Department o Computer Science and Engineering, Jadavpur University, Kolkata , India {kgparindamkar@gmail.com, debotosh@indiatimes.com, dipakkbasu@gmail.com, m.nasipuri@cse.jdvu.ac.in, mkundu@cse.jdvu.ac.in} Abstract This paper proposes a new technique or automatic ace recognition using integrated peaks o the Hough transormed signiicant blocks o the binary gradient image. In this approach irstly the gradient o an image is calculated and a threshold is set to obtain a binary gradient image, which is less sensitive to noise and illumination changes. Secondly, signiicant blocks are extracted rom the absolute gradient image, to extract pertinent inormation with the idea o dimension reduction. Finally the best itted Hough peaks are extracted rom the Hough transormed signiicant blocks or eicient ace recognition. Then these Hough peaks are concatenated together, which are used as eature in classiication process. The eiciency o the proposed method is demonstrated by the experiment on 1100 images rom the FRAVD ace database, 00 images rom the FERET database, where the images vary in pose, expression, illumination and scale and 400 images rom the ORL ace database, where the images slightly vary in pose. Our method has shown 93.3%, 88.5% and 99% recognition accuracy or the FRAVD, FERET and the ORL database respectively. Index Terms Face recognition, Feature extraction, Gradient image, signiicant blocks, Hough transormation, Hough peaks. I. INTRODUCTION Automated processing o ace images has been receiving increasing attention during the last ew decades. Image matching is a undamental aspect o many problems in computer vision, including object or scene recognition, solving or 3D structure rom multiple images, stereo correspondence, and motion tracking. Feature selection is one o the most important steps or detection and classiication problems. Good eatures should be discriminative, robust, easy to compute, and eicient. The Feature selection is speciically important or any ace recognition scheme. The more signiicant acial eatures such as outline o hair and ace, position o eyes, nose and mouth can be preserved by very small number coeicients. The raw pixel values o several image statistics such as colour, gradient and ilter responses is the simplest choice or image eatures, has been used or many years in computer vision, e.g., [1-3]. The underlying motivations or our approach originate rom the observation that humans achieve at least a basic level o categorization o aces at a glance even when images o very low resolution are used. Though human beings can detect and identiy aces with no eort, building an automated system that accomplishes such objectives is very challenging. The challenges are even more proound when there are large variations due to illumination conditions, change in pose, acial expression. O all image analysis o human ace, eature extraction is o immense importance. In order to construct the illumination cone, model based methods also require a set o training images or each subject. Furthermore, because o the complexity o light sources and illumination changes, how to compute and obtain physically implemented lower-dimensional subspaces basis images requires deep investigation. Edge relects the discontinuity o intensity distribution in an image. It contains contour, structure and shape inormation o objects in an image. Cognitive psychological studies indicated that human beings recognize line drawings as quickly and almost as accurately as gray-level pictures. Furthermore, edge is insensitive to illumination changes. In particular, we propose an approach based on ace similarity matching using Hough peaks [4] as the eature vector, and showed that this approach produces good recognition results even when less than 5% inormation o the original gray-scale image is retained ater selection o the Hough peaks. Alco in this paper a novel methodology applicable to ace matching and ast screening o large acial is introduced. The proposed shape comparison method operates on edge maps and derives holistic similarity s, yet, it does not require the solution o the correspondence problem. The use o edge images is important to introduce robustness against changes in illumination. There are three main contributions in this paper. Firstly, the gradient o several image statistics is computed or an image o interest, such as the image descriptor. Instead o the joint distribution o the image statistics, we use the gradient change as our edge points or Hough transormation, so the dimension becomes much smaller and hence the computational cost is reduces which is independent o the size o the image. Secondly, applying Hough transormation only on the ew blocks o the binary gradient image makes the computationally quite ast. As the block based acial eatures are compared locally, instead o using a general structure, allows to compare aces in terms o mouth, nose and other eatures in presence o occlusion. Finally we use the concept o

2 clustering to ind the centroid o all the Hough peaks obtained or each block and consider only the nearest two peaks instead o all the peaks. This paper describes image eatures that have many properties that make them suitable or matching dierent images o an object. The eatures are invariant to image scaling and rotation, and partially invariant to change in illumination [5]. The remainder o this paper is organized as ollows: Section II describes the extraction o most signiicant blocks rom the binary gradient image o the acial image. Section III details with Hough transormation and selection o the Hough peaks rom the selected blocks. Section IV deals with the similarity and classiication rule. In Section V and VI we assesses the perormance o the proposed method on the ace recognition task by applying it on the FERET [6], FRAVD [7] and the ORL ace [8] and inally by comparing with some o the most popular ace recognition schemes. A. Binary Gradient Image II. MAP OF AN IMAGE In this section the gradient image I G o an image (I) is calculated by taking summation over the absolute o the change o pixels in all the eight particular directions i.e., (north (N), northeast (NE), east (E), southeast (SE), south (S), southwest (SW), west (W) and northwest (NW)). The absolute binary gradient image I AG is calculated by considering the pixels o I G those are greater than the average o the mean and median o I G. The absolute gradient image is calculated as: I G abs abs x, y abs abs I x 1, y I x, y abs I x, y 1 I x, y I x 1, y I x, y abs I x, y 1 I x, y I x 1, y 1 I x, y abs I x 1, y 1 I x, y I x 1, y 1 I x, y abs I x 1, y 1 I x, y Ater calculating the gradient image, edge and Sobel operators [9] are used to calculate the threshold value, which in turn give a binary gradient image. The original image and its corresponding binary gradient image are shown in Fig. 1.(a) and Fig. 1. (b) respectively: Fig.1. (a) Fig.1. (b) Fig 1.(a) Original image, (b) Binary absolute gradient image The processed binary gradient mask images still shows lines o high contrast in the image. So by using linear structuring elements i.e. by dilation o the binary gradient image, these linear gaps are removed. This is very similar to the way by which human perceive the ace o other human being. From igure 1(b) it is clear that the white portions o the binary gradient image, where the change in contrast is very high, and are the region o interest except those boundaries i.e., we are interested to extract those regions that contains the important acial eatures like nose, mouth, eye etc. B. Extraction o Signiicant Blocks The step or extracting the signiicant blocks rom the image is given below. Blocks are generated randomly inside the absolute gradient image I G and are selected i; a) The count o white pixels in the block is greater than M % o the total pixels in the block, where M is the mean o the pixel values o the binary gradient image. b) I there is any overlap between a selected block and a new block, then that particular block is selected or which the count o white pixels is more. The extracted signiicant k th block or the i th training image is deined as B x, y, B x, y. The irst two i, j k k i, j k k components represent the starting location o the extracted block. These two components o eature are very important during matching (comparison) process. The remaining components are the elements o the block. d) This process o random block generation and extraction is repeated or X times (say X=500000), and the most signiicant blocks are selected rom the generated blocks in such a way that the blocks captures the important acial eatures like nose, mouth, eye brows, eyes etc and are in the range o 0-30% o the total blocks in the image. The proposed gradient based method reduces the number o edge pixels by as much as 70-80%. III. SIGNIFICANT BLOCK EXTRACTION A. Hough tranormation The Hough transorm is widely used or shape analysis in machine vision with its robustness to noise and high eiciency. In this paper, we ocus on the standard Hough transorm (SHT) [10], which is used to detect straight lines. The SHT uses the parametric representation o a line: x cos y sin, here ρ is the distance rom the origin to the line along a vector perpendicular to the line. θ is the angle o the perpendicular projection rom the origin to the line d in degrees clockwise rom the positive x-axis. The range o θ is The angle o the line itsel is 90 d clockwise with respect to the positive x-axis. where ρ is the length o the perpendicular o a line passing through (x,y) and subtending an angle θ with the x-axis. Since the point can be deined by a set o intersecting lines that subtend angles between 0 to 90, a sinusoidal locus is generated using x cos y sin, or every edge point. A straight line can be characterized by a set o

3 collinear edge points. So each o these edge points have one line in common, which implies that at one unique (ρ,θ) coordinate in Hough space the sinusoids o these points intersect. This is illustrated in Fig.. Fig. Representation o Hough transormation The SHT is a parameter space matrix whose rows and columns correspond to ρ and θ values respectively. The elements in the SHT represent accumulator cells. Peak values in the SHT represent potential lines in the input image. The steps or extracting o Hough peaks rom the signiicant blocks are given below. a. For each such block B i, i=1,, n Hough transormation is applied. b. Let the size o each Block be m m. Firstly consider the top m peaks in decreasing order o their value. c. Then ind the centroid o these m peaks using the K- means cluster algorithm [11], and select only those peaks which are nearest to the centroid as eatures. Thus we obtain a eature vector o size n, where n is the number o blocks. The proposed selection o peaks has been adjusted experimentally. Fig. 3(a), (b) and (c) show the selection o the signiicant block rom the binary gradient image, the corresponding block in the original and the selected Hough peaks rom the signiicant block rom the FERET dataset, respectively. F Fig. 3. (a) Fig. 3. (b) Fig. 3. (c) Fig 3. (a) Signiicant Block extracted rom the binary gradient image, (b) Same position block rom the original image (c) selected Hough peaks rom a signiicant block Thus rom each block the obtained Hough peaks are concatenated to obtain a eature vector o size n or n. This inal eature obtained rom the Hough peaks are points with high inormation content o the ace image. IV. SIMILARITY MEASURE AND CLASSIFICATION The dissimilarity between the i th testing and j th training images is d as: D B 1,, B i k 1 1 l B j i j i max I k, l., ik jl where B = number o blocks in the i th testing image, and i B = number o blocks in the j th training image. Here j I k, l is deined in such a way that blocks o the training image which are not close enough to a block o the test image in terms o location are discarded, and the extracted eatures (Hough peaks) o those training blocks which are in the neighborhood o the testing blocks are only compared or inding the similarity i.e, I k, l x x y y 1 i th1 k l k l, 0 otherwise where th1 is the threshold. In this experiment the threshold is taken as the block size i.e. the approximate radius o the region that contains the eye. This technique helps to avoid the matching o a eature o a block located around the eye with a point o a training acial image that is located around the mouth. Also, is the Chi-square [1] 1 dissimilarity between vector 1 and, and is deined as:, 1 m 1 1 m 1 m m m, where m is the number o eatures (Hough peaks) in the block. Finally, the i th testing image is classiied to the k th training image class i D i j min D i, j,. j V. EXPERIMENT The eectiveness o the proposed method has been successully tested on ace recognition using three, a) the whole ORL acial database, b) the FRAVD database containing 1100 rontal ace images corresponding to 100 subjects, c) The FERET database, containing 00 rontal ace images corresponding to 00 subjects, which are acquired under variable illumination and acial expression. The eectiveness o the method is shown in terms o both absolute perormance indices and comparative perormance against some popular ace recognition schemes such as the PCA, LDA [13, 14], Gabor wavelets (GW) [15], edge based eatures, direct Hough transormation, and sub peaks o the Hough transormation [16]. A. Experiment on the ORL database The whole ORL database is considered here. In the experiment each image is scaled to 9 11 with 56 gray levels.. Fig. 4 shows all samples o one individual. First two images o each individual are considered or training and the remaining images are used as testing samples.

4 (A) regular acial status; (B) +15 pose angle; (C) -15 pose angle; (D) +5 pose angle; (E) -5 pose angle; (F) +40 angle; (G) -40 pose angle; (H) +60 pose angle; (I) - 60 pose angle; (J) alternative expression; (K) dierent illumination. Fig. 4. Demonstration images o one subject rom the ORL database B. Experiment on the FRAVD database The FRAVD ace database, employed in the experiment, consists o 1100 colour ace images o 100 individuals, 11 images o each individual are taken, including rontal views o aces with dierent acial expressions, under dierent lighting conditions. All colour images are transormed into gray images and scaled to Fig. 5 shows all samples o one individual. The details o the images are as ollows: (A) regular acial status; (B) and (C) are images with a 15 turn with respect to the camera axis; (D) and (E) are images with a 30 turn with respect to the camera axis; (F) and (G) are images with gestures; (H) and (I) are images with occluded ace eatures; (J) and (K) are images with change o illumination. Fig 5. Demonstration images o one subject rom the FRAVD database C. Experiment on the FERET database The FERET database, employed in the experiment here, contains,00 acial images corresponding to 00 individuals with each individual contributing 11 images. The images in this database were captured under various illuminations that display, a variety o acial expressions and poses. As the images include the background and the body chest region, so each image is cropped to exclude those, and then scaled to Fig. 6 shows all samples o one individual. The details o the images are as ollows: Fig. 6. Demonstration images o an individual rom the FERET database D.Speciicity and Sensitivity or the FRAVD,FERET and the ORL dataset: To the sensitivity and speciicity [17] the dataset rom the FRAVD, & FERET database is prepared in the ollowing manner. In the FRAVD database or each individual a single class is constituted with 18 images in it. Thus, a total 100 class is obtained rom the dataset o 1100 images o 100 individuals. Out o the 18 images in each class, 11 images are o a particular individual, and 7 images are o other individuals taken by permutation. Similarly 00 classes are obtained or the FERET dataset, each class having 18 images in it. Out o the 18 images in each class, 11 images are o a particular individual, and 7 images are o other individuals taken by permutation. For ORL database 40 classes are constructed with 15 images in each class. Out o the 15 images in each class, 10 images are o a particular individual, and 5 images are o other individuals. Using these datasets the true positive (T P ); alse positive (F P ); true negative (T N ); alse negative (F N ); are d. For all the datasets, the irst images (A-B), o a particular individual are selected as training samples and the remaining images o that particular individual are used as positive testing samples. The negative testing is done using the images o the other individuals. Table I: Speciicity and Sensitivity o the FERET & FRAVD dataset: Total no. o classes=100, Total no. o images= 1800 FRAVD test FRAVD Individual belonging to a particular class Using irst images o an individual as training images Positive Negative Positive T P =840 F P =7 Negative F N =60 T N =693 Sensitivity = T P / (T P + F N) 93.33% Speciicity =T N / (F P + T N) 99.3%

5 Total no. o classes=00, Total no. o images= 3600 FERET test ORL test FERET Using irst images o an individual as training images Positive Negative Positive T P = 1593 F P =14 Negative F N =07 T N =1386 Sensitivity = T P (T P + F N) 88.5% Speciicity =T N / (F P + T N)=99% Total no. o classes=40, Total no. o images= 600 Using irst images o an individual as training images Positive T P =316 F P =0 Negative F N =4 T N =00 Sensitivity = T P (T P + F N) 99% Speciicity =T N / (F P + T N)=100% Thus or FRAVD database considering the irst images (A-B) o a particular individual or training the achieved rates are: False positive rate = FP / (FP + TN) = 100% Speciicity =.7% False negative rate = FN / (TP + FN) = 100% Sensitivity=6.67% Accuracy = (T P +T N )/(T P +T N +F P +F N ) 96.3%. For FERET database considering the irst images (A-B) o a particular individual or training the achieved rates are: False positive rate = FP / (FP + TN) = 100% Speciicity =1% False negative rate = FN/(TP+ FN) = 100% Sensitivity=11.5% Accuracy = (T P +T N )/(T P +T N +F P +F N ) = 93.75%. For ORL database considering the irst images (A-B) o a particular individual or training the achieved rates are: False positive rate = FP / (FP + TN) = 100% Speciicity =0% False negative rate = FN/(TP+ FN) = 100% Sensitivity=1% Accuracy = (T P +T N )/(T P +T N +F P +F N ) = 99.5%. VI. RESULTS Experimental results indicate that a) the extracted Hough peaks as eatures, rom the proposed signiicant blocks o the absolute gradient image integrated together used or classiication using the distance similarity achieved a veriication rate which is as good as any previously used methods like PCA, LDA, Gabor wavelets or ace recognition.it is also seen that signiicant block based binary gradient based eature extraction technique achieves the best accuracy, when peaks are selected rom the Hough transormed blocks.; b) During simulations, it is observed that the locations o signiicant blocks, ound rom the absolute binary gradient image o the ace image, can give small deviations between dierent conditions (expression, illumination, having glasses or not, rotation, etc.), or the same individual. Thereore, an exact ment o corresponding distances is not possible unlike the geometrical eature based methods; c) The computational time is signiicantly reduced by using only ew signiicant blocks o the image, although the accuracy is not compromised; and inally d) selected o only those peaks rom the Hough transormed block rom the accumulation o peaks which are nearest to the centroid o the available peaks, and hence reduce the size o the eature vector. Extensive experiments indicate that the proposed method has achieved a much better perormance than the previous variations o Hough transorm. Further it also decreases the eect o occluded eatures. Table II shows the upper bound perormances on the FRAVD and ORL acial, which relects that our proposed method achieves higher perormance results. Thus the proposed algorithm deals with two o these problems, namely occlusion and illumination changes. Experimental results on all the three taking irst two images (A-B) as training ace and the remaining images as testing images are shown below. Table II. Perormance results o well known ace recognition algorithms together with the proposed method on ORL and FRAVD respectively with the use o Databases FRAVD, FRAVD, FRAVD, FRAVD, FRAVD, FRAVD, FRAVD, FRAVD, FRAVD, FRAVD, FRAVD, Experiment conducted or eature Selection Signiicant Blocks o the Binary Gradient Image + Hough Transormation+ Selection o the nearest to the centroid peaks. Signiicant Blocks o the Binary Gradient Image + Hough Transormation+ Selection o the all the m peaks, [m= Block size]. Selecting all the Blocks o the Binary Gradient Image + Hough Transormation+ Selection o the nearest to the centroid peaks. Selecting all the Blocks o the Binary Gradient Image + Hough Transormation + Selection o the all the m peaks, [m=block size]. Binary Gradient Image (without block breaking)+ Hough Transormation + Selection o all the available peaks. Signiicant Blocks o the Binary Gradient Image Sub Peaks o Hough transormation All the edge points o the Binary Gradient Image similarity. Similarity Measure Gabor Wavelets LDA PCA Highest Recognition Accuracy (%) 88.5, 93.3, 99 85,88.8, , 88, 95 84,88, ,80.5, 95 80, 83.8, , 91, ,86, ,85.4, ,86.7, ,8.8,97.5

6 CONCLUSION The proposed technique as presented here obtains the absolute gradient o the original image in eight directions. From these gradient images only the inormative signiicant blocks are extracted with our extraction technique, and are used, thus the number o edge points are drastically reduced which alleviate the computational and storage load, or real time applications. As only the signiicant blocks are used instead o the whole image, so in this approach the acial eatures are compared locally instead o a general structure, and hence allow us to make decision rom the dierent parts o a ace. As such it perorms better in presence o occlusions. Again by selection o only those two peaks o the Hough transormed signiicant blocks that are nearest to the centroid o the available peaks urther reduces the dimension o the eature vector and also enhances ace recognition in presence o illumination and expression changes as a property o Hough transormation. Transaction on Pattern Analysis and Machine Intelligence, Vol. 4, No. 7, JulyY 00. elligenc [1] [13] M. Turk, A. Pentland, Eigenaces or Recognition, Journal o Cognitive Neuroscience, vol. 3, pp X, [14] A. Pentland, B. Moghaddam, T. Starner, View Based and Modular Eigenspaces or Face Recognition, In Proceedings o Computer Vision and Pattern Recognition, pp , , IEEE Computer Society, Seattle, USA, June [15] C. Liu, H. Wechsler, Independent Component Analysis o Gabor Features or Face Recognition, IEEE Transactions on Neural Networks, vol. 14, no. 4, pp , July [16] Xiaohui Chen & Xiaojun Liu A sub peak tracker based Hough transorm or accurate and robust linear edge extraction 010 International Conerence on Electrical and Control Engineering. [17] speciicityaccuracy -and relationship between them. ACKNOWLEDGEMENT Authors are thankul to a major project entitled Design and Development o Facial Thermogram Technology or Biometric Security System, unded by University Grants Commission (UGC),India and DST-PURSE Programme and CMATER and SRUVM project at Department o Computer Science and Engineering, Jadavpur University, India or providing necessary inrastructure to conduct experiments relating to this work. Dr. D.K. Basu would also like to thank the AICTE, New Delhi, INDIA or providing him an Emeritus Fellowship (F.No.1-51/RID/EF (13)/ ). REFERENCES [1] R. Mar ee, P. Geurts, J. Piater, and L. Wehenkel, Random subwindows or robust image classiication, in Proc. IEEE Con. on Computer Vision and Pattern Recognition, San Diego, CA. vol. 1, pp , 005. [] F. Porikli, Integral histogram: A ast way to extract histograms in Cartesian spaces, in Proc. IEEE Con. on Computer Vision and Pattern Recognition, San Diego, CA. vol. 1, pp , 005. [3] A. Ramisa, S. Vasudevan, D. Scharamuzza, R. M antaras, and R. Siegwart, A tale o two object recognition methods or mobile robots, Artiicial Intelligence Research Institute (IIIA-CSIC), Campus UAB, Bellaterra, Spain. [4] [5] D. Lowe, Distinctive image eatures rom scale-invariant keypoints, Intl. J. o Comp. Vision vol. 60, pp , 004. [6] The Face recognition technology (FERET) database, available at: [7] Face Recognition and Artiicial Vision group FRAVD ace database available at [8] The Oracle Research Laboratory Face Database o Faces, [9] H. B. Kekre, Ms. Saylee M. Gharge Image Segmentation using Extended Edge Operator or Mammographic Images,International Journal on Computer Science and Engineering Vol. 0, No. 04, 010, [10] [11] Tapas Kanungo, David M. Mount, Nathan S. Netanyahu, Christine D. Piatko, Ruth Silverman and Angela Y. Wu, An Eicient k- Means Clustering Algorithm: Analysis and Implementation IEEE