ROBUST MULTI GRADIENT ENTROPY METHOD FOR FACE RECOGNITION SYSTEM FOR LOW CONTRAST NOISY IMAGES

Transcription

1 ROBUST MULTI GRADIENT ENTROPY METHOD FOR FACE RECOGNITION SYSTEM FOR LOW CONTRAST NOISY IMAGES C. Naga Raju 1, P.Prathap Naidu 2, R. Pradeep Kumar Reddy 3, G. Sravana Kumari 4 1 Associate Professor, CSE Dept, YSR Engg College of YVU 2 Asst. Professor, CSE Dept, RGM Engg College 3 Asst. Professor, CSE Dept, YSR Engg College. 4 M.Tech In CSE RGM Engg College Abstract one of the most important challenges for practical face recognition systems is making recognition more reliable under uncontrolled lighting and noisy conditions. in this paper we made three main contributions: (i) we presented a simple and efficient preprocessing method that eliminates the noise and most of the effects of changing illumination by preserving the essential appearance details that are needed for recognition;(ii) we introduced a generalized Local ternary Pattern (LTP) for local feature description that is more discriminate and less sensitive to noise in uniform regions, and (iii) we further improved robustness by applying entropy method for feature extraction from face images. The resulting method provides state-ofthe-art performance on three data sets that are widely used for testing recognition under difficult illumination conditions. Keywords:Multi_Gradient, Similarity Metric. LTP,Entropy,Feature sets, I. INTRODUCTION Face recognition has received a great deal of attention from the scientific and industrial communities over the past several decades owing to its wide range of applications in information security and access control, law enforcement,surveillance and more generally image understanding[1].most of these methods were initially developed with face images collected under relatively well controlled conditions and in practice they have difficulty in dealing with the range of appearance variationsthat commonly occur in unconstrained natural images due to illumination, pose, facial expression, ageing, partial occlusions, etc. This paper focuses mainly on the issue of robustness to lighting variations. For example, a face verification system for a portable device should be able to verify a client atany time (day or night) and in any place (indoors or outdoors).unfortunately, facial appearance depends strongly on the ambient lighting and as emphasized by the recent FRVT and FRGC trials [2,3]this remains one of the major challenges for current face recognition systems.we will investigate several aspects of this framework:1) The relationship between image normalization and feature sets. Normalization is known to improve theperformance of simple subspace methods (e.g. PCA) orclassifiers (e.g. nearest neighbors) based on image pixel representations [4], but its influence on more sophisticated feature sets has not received the attention that it deserves. For example, for Histogramof Oriented Gradient features combining normalizationand robust features is useful in [5], while histogramequalization has essentially no effect on LBP descriptors[6,7], and in some cases preprocessing actually hurtsperformance [8,9,10] presumably because it removes toomuch useful information. However in complex tasks such as face recognition, it is often the case that no single class of features is rich enough to capture all of the available information. Finding and combining complementary feature sets has thus become an active research topic, with successful applications in many challenging tasks including handwritten character recognition [11] and face recognition [12]. Here we show that combining three of the most successful local face representations, multiple gradient method and Local Ternary Patterns (LTP), and entropy based method, gives considerably better performance than either alone. The three feature sets are complimentary in the sense that LTP captures small appearance details while entropy preserves facial shape over a broader range of scales. 2. PROPOSED METHOD 2.1. Multi gradientmethod This algorithm is computed based on the notion of regional maxima, regional minima, most significant value in the region and regional sorted Meddle values and uses different gradient algorithms constructed based on regional maxima, regional minima and regional sorted Meddle values for reconstruction. A pixel p of I isa local maximum for grid G if and only if its valuei (p) is greater or equal to that of any of itsneighbors. A pixel p of I is a local minimum forgrid G if and only if its value I (p) is less than orequal to that of any of its neighbors. A pixel p of Iis a local median for grid G if and only if its value I(p) is in N/2 position of the sorted grid G of itsneighbors. A pixel p of I is a most significant forgrid G if and only if it has more neighbors than anypixel in the grid. Next we Volume 2, Issue 3 May June 2013 Page 193

2 compute gradient ofmiddle and minima (Gmmin), gradient of middleand maxima (Gmmax), gradient of most significantand minima (Gsmin) and gradient of most significant and maxima (Gsmax). By using thesegradients we can generate single value at p(x,y)location of resultant Image Algorithm. Step1: Read the color image and convert it togray-scale. Step2: Find region minima, region maxima,significant value and sorted middle Value Step3: compute Gmmin, Gmmax Step4: if Gmmin<0 and Gmmax>0 thenr=significant pixel and gotostep6 ElseCompute Gsmin, Gsmax Step5: if Gsmin> 0 and Gsmax<0 then Result=significant pixel Else Result =sorted middle pixel Step6: repeat step2 to step6 by convolution 2.2.Local Ternary Patterns (LTP) The LTP has 3-valued codes, LTP, in which gray-levels in a zone of width around are quantized to zero, ones above this are quantized to 1 and ones below it to 1, i.e., the indicator is replaced with a 3-valued function Here t is a user-specified threshold for LTP so these codes are more resistant to noise, but no longer strictly invariant to gray-level transformations. The LTP encoding procedure is illustrated in fig7. Fig1: Illustration of the basic LTP Operator When using LTP for visual matching we could use 3- valued codes, but the uniform pattern argument also applies in the ternary case. For simplicity, the experiments below use a coding scheme that splits each ternary pattern into its positive and negative halves as illustrated in Fig. 2, subsequently treating these as two separate channels of LBP descriptors for which separate histograms and similarity metrics are computed, combining the results only at the end of the computation. LTP s bear some similarity to the texture spectrum (TS) technique. However TS did not include preprocessing, thresholding, local histograms or uniform pattern based dimensionality reduction and it was not tested on faces. Fig2: Splitting an LTP Code into positive and negative LBP Codes 2.3. Entropy method We further improve robustness by adding entropy feature extraction.in this method the optimal threshold value exists at the valley of the two peaks or at the bottom rim of a single peak. The valley in the histogram that separates the facial features from the image, its probability of occurrence is small in gray level histogram. Because of the optical threshold should near the cross where the facial features and the image back ground intersect. The probability of occurrence at the threshold value should divide into two parts. Its first half belongs to background and second half belongs to image facial features. Then we apply a new weight M to themethod. t = (P1*D1+ P2*D2)* M Where D1=(σ1 2 (o)- σt 2 (t))2 D2=(σ2 2 (b)- σt 2 (t))2 M=(1-PT(t)/2) Using this method we can make sure that the result threshold value resides at the valley or at the bottom of the right rim of single peak. It s also maximizes group variance and ensures that both the variance of the facial features and that of the image background keep away from the variance of the whole image. Smaller the p(t)/2, larger will be the weight Similarity Metric The LTP method for face recognition that divides the face into a regular grid of cells and histograms the uniform LTP s within each cell, finally using nearest neighbor classification in the Image histogram distance for recognition: Here p; q are image region descriptors (histogram vectors), respectively. This method gave excellent results on the FERET dataset. However subdividing the face into a regular grid seems somewhat arbitrary: the cells are not necessarily well aligned with facial features, and the partitioning is likely to cause both aliasing and loss of spatial resolution. Given that the overall goal of coding is to provide illumination- and outlier-robust visual correspondence with some leeway for small spatial deviations due to misalignment, it seems more appropriate to use a Hausdorff-distance-like similarity metric that takes each LTP pixel code in image X and tests whether a similar code appears at a nearby position in image Y, with a weighting that decreases smoothly with image distance. Such a scheme should be able to achieve discriminant appearance based image matching with a well-controllable degree of spatial looseness. 3. Experimental results LBPs have proven to be highly discriminative features for image classification and they are resistant to lighting effects in the sense that they are invariant to monotonic gray-level transformations. However because they threshold at exactly the value of the central pixel they tend to be sensitive to noise, particularly in near-uniform Volume 2, Issue 3 May June 2013 Page 194

3 image regions, and to smooth weak illumination gradients. Many facial regions are relatively uniform and it is legitimate to investigate whether the robustness of the features can be improved in these regions. Fig1 shows the output of the proposed method at each level. The original Image has salt & pepper noise and it has low contrast, after applying multi gradient step noise has been eliminated and contrast has been improved in the next step applying the LTP operator feature have been extracted with false features in the next step entropy has been improved overall features of the images. Finally nearest neighborhood classifier applied for face detection. This method is applied for different imagesand results have been shown in the paper through case1 to case4. Q1,Q4 and Q6 are the query Images compared with database images like a1to a9,p1to p9 s1to s9 and n1 to n9 and produced tables of values and graphs.the tables of values, graphs and images show the matching rate of face. 3.1Stepwise outputs of proposed method Graph1 Fig3: Database images (first row), LBP (second row), proposed(third row) Table2 Fig1:a)Original image b)multi gradient c)ltp image d)entropy image 3.2 Test cases: Graph2 Fig2.Database images (first row), LBP (second row), proposed (third row) Table1 Fig4: Database images (first row),lbp (second row),proposed(third row) Table3 Volume 2, Issue 3 May June 2013 Page 195

4 means that only one person will be facing the camera at any one time. In future extensions of the system, the intelligent sales assistant will be able to analyze the facial expressions of groups, and make a decision based on a number of measures including averaging the affect detected. Graph3 Fig5: Database images (first row),lbp (second row),proposed(third row) Table4 Graph4 4. CONCLUSION AND FUTURE SCOPE The primary question that springs to mind when discussing facial expression recognition is the accuracy and repeatability of such a system. The result of the proposed algorithm improves performance substantially with respect to the original LBP-based algorithm when used in relatively unconstrained face datasets. The proposed method also outperforms the original LBP algorithm even when faces are frontal and well aligned,though by a smaller margin. This improvement may be attributed to the flexible spatial matching scheme and the use of the image-to-class distance,which makes a better use of the training data than the image-toimage distance.the system described in this paper assumes that it deals with one by one basis of faces. This 5. References [1] W. Zhao, R. Chellappa, P. J. Phillips, and A. Rosenfeld, Face recognition: A literature survey, ACM Computing Surveys,vol. 34, no. 4, pp , [2] L.Sirovich and M. Kirby, Low dimensional procedure for the characterization of human faces, J. Optical Society of America, vol. 4, no. 3, pp , [3] P. J. Phillips, P. J. Flynn, W. T. Scruggs, K. W. Bowyer, J. Chang, K. Hoffman, J. Marques, J. Min, and W. J. Worek, Overview of the face recognition grand challenge, in CVPR, San Diego, CA, 2005, pp [4] H. Wang, S. Li, and Y. Wang, Face recognition under varying lighting conditions using selfquotient image, in IEEE Int. Conf. Automatic Face & Gesture Recognition, 2004, pp [5] N. Dalal and B. Triggs, Histograms of oriented gradients for human detection, in CVPR, Washington, DC, USA, 2005, pp [6 Face description with local binary patterns: Application to face recognition, IEEE Trans. Pattern Analysis & Machine Intelligence, vol. 28, no. 12, [7] W. Gao, B. Cao, S. Shan, X. Chen, D. Zhou, X. Zhang, and D. Zhao, The CAS-PEAL large-scale chinese face database and baseline evaluations, IEEE Trans. Systems, Man and Cybernetics, Part A, vol. 38, no. 1, pp , [8] T. Zhang, D. Tao, X. Li, and J. Yang, Patch alignment for dimensionality reduction, IEEE Trans. Knowledge & Data Engineering, vol. 21, no. 9, pp , [9] H. Chen, P. Belhumeur, and D. Jacobs, In search of illumination invariants, in CVPR, 2000, pp. I: [10] T. Chen, W. Yin, X. Zhou, D. Comaniciu, and T. Huang, Total variation models for variable lighting face recognition, IEEE Trans. Pattern Analysis & Machine Intelligence, vol. 28, no. 9, pp , [11] Y. S. Huang and C. Y. Suen, A method of combining multiple experts for the recognition of unconstrained handwritten numerals, IEEE Trans. Pattern Analysis & Machine Intelligence, vol. 17, no. 1, pp , [12] C. Liu and H. Wechsler, A shape- and texture-based enhanced fisher classifier for face recognition, IEEE Trans. Image Processing, vol. 10, no. 4, pp , Volume 2, Issue 3 May June 2013 Page 196

5 Author Dr. C. Naga Rajureceived his B.Tech degree in Computer Science from J.N.T.UniversityAnantapur, M.Tech degree in Computer Science from J.N.T.University Hyderabad and Ph.D on digital Image processing from J.N.T.University Hyderabad. Currently, he is working as Associate professor in CSE Department at Y S R Engineering College of Yogi Vemana University, Produttur, Kadapa. He has got 17 years of teaching experience. He has published thirty six research papers in various national and international journals and about thirty four research papers in various national and international conferences. He has attended twenty seminars and workshops. He is member of various professional societies of IEEE, ISTE and CSI. P.Prathap Naidu received his AMIE Degree in Computer Science frominstitute of Engineers(india), Kolkata,M.Tech Degree in Computer Science from J.N.T.University hyderabad. He has six years of teaching experience.at present he is working as Assistant Professor in CSE at Rajeev Gandhi Memorial College of Engineering & Technology,Nandyal, A.P, India.he has attended three work shops and four conference. He is a member of professional society of IE. R. Pradeep Kumar Reddy received his B.Tech. Degree in Computer Science and Engineering from Bellary Engineering College, Bellary(VTU).M.Tech. Degree in Computer Science and Engineering at S.R.M University, Chenai. Currently He is working as Assistant Professor in the Department of CSE at YSR Engineering College of Yogivemana University,Proddatur. He has got 9 years of Teaching experience. He has attended 5 workshops. He is a member of ISTE. G. Sravana Kumari received her B.Tech Degree in Computer Science and Engineering from Annamacharya institute of technology & sciences,tirupati. At present she is pursuing her M.Tech Degree in Computer Science and Engineering at Rajeev Gandhi Memorial College of Engineering & Technology, Nandyal, A.P, India. She is the topper in the class. Volume 2, Issue 3 May June 2013 Page 197