Face recognition system using HGPP
Over the last decade, human face recognition has attracted significant attention because of its wide range of applications, such as criminal identification, credit card verification, security system, scene surveillance, entertainments etc., In these applications various face recognition techniques are used and these techniques depend on the intensity of an image, variations in the image such as shape, lighting etc., To eliminate extrinsic factors, various features extraction and selection methods are used. One method is Histogram of Gabor phase pattern (HGPP). In this method the quadrant bit codes are first extracted from face, based on the Gabor transformation and Histogram techniques. The features of HGPP lie in two aspects, they are: i) HGPP can describe the general face images robustly without training procedure, ii) Encodes the Gabor phase Information, instead of Gabor magnitude information. This method uses LGPP and GGPP to encode the phase variations, as this uses high dimensional histogram features, this method lacks performance. Computational complexity arises due to the huge volume of data processing. Also the time taken to process the data is increased. In this design, the above stated problems are rectified using adaptive binning method which increases the overall efficiency of the face recognition system.
Face recognition, Gabor Wavelets, Local Gabor Phase pattern (LGPP), Global Gabor Phase Pattern (GGPP), Adaptive Binning, Spatial Histograms, HGPP.
Face recognition is the fledgling of the most coveted field-Image Processing. Though there are algorithms to accomplish this novel task which underpins access to sensitive data and entry into highly secure places and workplace otherwise, there is an imminent crave to develop a highly efficient, guiltiness method which wraps in itself the elixir to overcome the innate technical in competencies found in existing algorithms, upon which we dwell. Face recognition is co-founded with difficulties due to false alarms given off by ineffective recognition strategies. Face recognition has been an active research topic over the past few years due to its scientific challenges and potential applications. The Gabor wavelets approach appears to be quite perspective and has several advantages such as invariance to some degree with respect to homogenous illumination changes, small changes in head poise and robustness against facial hair, image noise [1,2,3,4,5]. Experimental results show that the proposed method performs better than traditional approaches in terms of both efficiency and accuracy. Gabor function being dual based and non-orthogonal is computationally resource intensive.
Challenges Associated with Face Verification
Face verification and recognition is a challenging problem due to variations in pose, illumination, and expression. Techniques that can provide effective feature representation with enhanced discriminability are crucial . Face recognition has become one of the most active research areas; it plays an important role in many application areas, such as human machine interaction, authentication and surveillance. However the wide range variations of human face, due to pose, illumination, and expressions, result in a highly complex distribution and deteriatorate the verification rate. It seems impractical to collect sufficient prototype images covering all the possible variations. Therefore, how to construct a small size training face verifier and robust to environmental variations is a challenging research issue.
The figure 1 shows the entire block design of the new system with new methodology.
Gabor feature has been recognized as one of the best representations for face recognition. Traditionally, only the magnitudes of the Gabor coefficients are thought to be valuable for face recognition, and the phases of Gabor features are deemed to be useless and always discarded directly by almost all researchers in face recognition community [7,8,9]. By encoding Gabor phases through Local Binary Pattern (LBP) and the spatial histograms achieved encourages recognition rate comparable with that of Gabor magnitude based methods. And it is also shown that the Gabor phases are quite compensatory to the magnitude information, since higher classification accuracy is achieved by combining Gabor phases and magnitudes. All these observations suggest that more attention should be paid to Gabor phases for face recognition. Among various wavelet bases, Gabor functions provide the optimized resolution in both the spatial and frequency domains. Gabor wavelets seem to be the optimal basis to extract local features for pattern recognition for several reasons:
- Biological motivation: The shapes of Gabor wavelets are similar to the receptive fields of simple cells in the primary visual cortex.
- Mathematical motivation: The Gabor wavelets are optimal for measuring local spatial frequencies.
- Empirical motivation: Gabor wavelets have been found to yield distortion tolerant feature spaces for other pattern recognition tasks, including texture segmentation, handwritten numeral recognition and fingerprint recognition.
Gabor Wavelets Functions
Gabor wavelet are obtained by using eq-1
D & E are calculated using L2 norms. L2 norm is defined as the square root of sum of squares of individual components.
In the above eq(1), u refers to the orientation and v refers to frequency, fmax is a constant. The above equation can be simplified further as shown below:
The real and imaginary part of the Gabor wavelets is applied with the Daugman's Method proposed by Daugman for demodulation. Daugman proposed this method for iris recognition. When the output of Gabor Wavelets is demodulated, each pixel in the resultant image will be encoded to two bits. This method is essential to split the Gabor Wavelets Pattern to GGPP and LGPP . Daugman's Functions:
Formation of GGPP
GGPP scheme computes one binary string for each pixel by concatenating the real (or imaginary) bit codes of different orientations for a given frequency. Formally, the GGPP value, GGPPv(Zo), for the frequency v at the position Zo in a given image is formulated as the combination of Daugmans Values. By using this encoding method, a decimal numbers for each pixel corresponding to the real GGPPs is obtained. GGPP scheme computes one binary string for each pixel by concatenating the (real or imaginary) bit codes of different orientations for a given frequency.
The above equation gives the real GGPP. In this approach there are eight orientations, which happen to form a byte of 28 to represent 256 different orientation modes. These modes can be easily computed by the following equations:
Formation of LGPP
LGPP is the further encoding of local variations for each pixel. LGPP actually encodes the sign difference of the central pixel from its neighbors. LGPP reveal the spots and flat area for the given images. Formally, for each orientation u and frequency v, the real LGPP value at the position Zo is computed using local XOR pattern (LXP) operator. The local variation for each pixel obtained is LGPP. LGPP actually encodes the sign difference of the central pixel from its neighbors. LGPP reveals the spots and flat area for the given images.
where Zi, i = 1,2,...8 are the eight neighbors around Z0, and XOR denotes the bit exclusive or operator.
Generation of HGPP Patterns
After using the Spatial Histograms and Adaptive Binning, a new pattern called HGPP is developed which has reduced data size, less complexity and most of the unwanted data are removed and hence the performance is increased. Using this HGPP Patterns the test images for the specific database will be checked and verified.
Adaptive Binning Algorithm
Histograms are used in image retrieval systems to represent the distributions of colors in images.The histograms adapt to images which represent their color distributions more efficiently than histograms with fixed binnings. Adaptive histograms produce good performance, in terms of accuracy, less number of bins and efficient computation when compared to existing methods for retrieval, classification, and clustering tasks. There are two general methods of generating histograms: fixed binning and adaptive binning. Adaptive binning is similar to color space clustering in that k-means clustering or its variant is used to induce the bins. However, the clustering algorithm is applied to the colors in an image instead of the colors in an entire color space. Therefore, adaptive binning produces different bins for different images. The adaptive binning algorithm is applicable to a wide range of data, from observations or numerical simulations, and is not limited to two dimensional data [12,13,14]. Adaptive binning is the simplest case of the algorithm. It attempts to adaptively bin a single image based on the number of photons in each region. The basic method is to bin pixels in two dimensions by a factor of two, until the fractional Poisson error of the count in each bin becomes less than or equal to a threshold value. When the error is below this value, those pixels are not binned any further. The algorithm, in detail, is as follows:
Each pixel in the image is put into its own 'bin', the term used for a collection of pixels. A pixel here means one of the individual picture elements which form the input image. Essentially the image is initially divided into imaginary 1 × 1 pixel bins.
If there are ni pixels in bin i, the total count in the bin is ci, and the background per pixel is b, the net count in the bin is simply defined by
The fractional error on the net count in the bin is
This is also the error on the average count in each pixel.
If the fractional error is less than or equal to a threshold value, then the pixels in the output image which correspond to the pixels in the input bin are set to the average mean count, si/n. The fractional error of the net count in the bin is also stored in the pixels in an 'error image'. Additionally the bin is marked as having been processed.
Each bin is merged into its neighboring three bins, to make new bins containing 2 × 2 of the previous bins. The four bins with the lowest x and y coordinates (lowest declination and highest right-ascension) are merged, as is each consecutive set of four bins. Any bins which have already been processed are ignored in the merging. It is useful to remember a bin as a list of pixels. Pixels which have already been set in the output image are ignored in future iterations.
The process is repeated from (ii) until there is only a single bin remaining.
A 'bin-map' is also produced by the algorithm, giving identification number for each processed bin in terms of the pixels which it contains. Using the bin-map, any image of that size can be binned using the same bins.
Object representation and feature extraction are essential to object detection. Specially, objects are modeled by their spatial histograms over local patches and class specific features are extracted. Spatial histograms consist of marginal distributions of an image over local patches; they can preserver texture and shape information of an object simultaneously . In this approach, a sub window containing a grey sample image with a certain size. GGPP and LGPP are used to preprocess sample images. GGPP & LGPP are relatively new and simple texture model and it has been proved to be a very powerful feature in classification images. GGPP & LGPP are invariant against any monotonic transformation of the gray scale. The basic GGPP & LGPP operator uses neighbourhood intensities to calculate the region central pixel value . The 3 x 3 neighbourhood pixels are signed by the value of center pixel:
The signs of the eight differences are encoded into an 8-bit number to obtain LGPP value of the center pixel:
For any sample image, histogram-based pattern representation is computed as follows. First, variance normalization on the gray image to compensate the effect of different lighting conditions are applied, then use basic global or local binary pattern operator to transform the image into an GGPP or LGPP image, and finally compute histogram of an image as representation. It is easy to prove that histogram, a global representation of image pattern, is invariant to translation and rotation. However, histogram is not adequate, since it does not encode spatial distribution of objects. For some non-object images and object images, their histograms can be very similar or even identical, making histogram not sufficient.
After using the Spatial Histograms and Adaptive Binning, a new pattern called HGPP is developed, which has reduced data size, less complexity and most of the unwanted data are removed and hence the performance is increased.
FERET Database is a collection of Face data with Different Orientations and properties. It is used to verify and validate the current Face Recognition system design to ensure the efficiency of the system to overcome the demerits of the existing system. This database consisting of 14000 thousand Face Images.
In this paper a new design is discussed, and the still images are separated in the Gabor Wavelets, Which obtained the Gabor Phases for the real and imaginary parts. The output of the Gabor Wavelets are given as input to the Daugman's Method, which encodes each pixel of an image, then the output of the Daugman's Method is taken as input, to further split the Global Gabor Phase Patterns and Local Gabor Phase Patterns. Then the Adaptive binning method is applied to the output obtained from the two Gabor Phase Patterns (GGPP & LGPP). After applying the adaptive binning technique, the size of the data can be reduced and the result is applied to generate a Histogram Gabor phase pattern. The results obtained with new design is compared with the existing method and checked with FERET Database.
Input Image Database
The database is divided into 3 sizes of 64x64, 88x88 and 128x128. In each size of image, we categorized the images into 10 parts as shown below:
Input Images (Sample)
In the Fig-3, shows the nine sample images of FERET data base of three different sizes.
Output - Gabor Wavelets
The Gabor Wavelets obtained for a sample images are shown in the figure 4 & 5. The below given images are generated keeping Frequency (v=5) and Orientation (u-8). So, totally 40 images are obtained.
Consolidated Result for an image
The Table-2 consists the sample Results obtained with the entire new system for an image. The HGPP pattern for the both the existing method and the new design are shown in figure 6 & 7 for the given input images.
The recognition rates for different sizes of images are tabulated in the Table 3 and the results are plotted as shown in figure 8. To further validate the effectiveness of HGPP, we compare our results with available methods such as Feature Extraction, Eigen Face and HGPP (i.e. without using Adaptive Binning Method). The results clearly indicate that the proposed HGPP method outperform all the other methods, especially on the Dup I, and Dup II probe set. Experiment results of this comparison evidently illustrate that the proposed HGPP method achieves the best results on our database. Since the face images in our database probe sets contain several source of variations such as expression, lighting, and aging, these comparisons illustrate that HGPP is impressively robust to these extrinsic imaging conditions, Gabor features can exhibit the spatial frequency (scale), spatial locality and orientation selectivity properties corresponding to Gabor wavelets. Adaptive Binning is a kind of quantification of Gabor feature, contributes to the robustness of HGPP.
In the Table 4 and 5 shows the recognition rate comparisons with other state of art results tests on our database probe sets.
The tabulated results are plotted (Figure 9 & 10) and highlights the impact of the results with different types of data. The new system achieves best results as per the recognition rate. The computational complexity due to the huge volume of database is reduced and gives opportunity to extend the database size. The future work will be in the maintainable and the linking databases of different features to make unique system for face recognition. The classification using the advanced level binning technique will improve the Dup I and Dup II results.
- Baochang Zhang, Shiguang Shan, Xilin Chen & Wen Gao, "Histogram of Gabor Phase Patterns (HGPP). A Novel Object Representation Approach for Face Recognition", IEEE Transactions on Image Processing, vol 16, No.1, pp 57-68, 2007.
- Xudong Xie, Kin-Man Lam, "Gabor Based Kernel PCA With Doubly Nonlinear Mapping for Face Recognition With a Single Face Image",IEEE Transactions on Image Processing, vol 15, No.9. pp. 2481-2492, 2006.
- W.Zhao, R.Chellappa, P.J.Phillips and A. Repsenfeld, "Face Recognition: A Literature survey", ACM Computing Survey., vol, 35, no. 4 pp:399-458,2003.
- R. Chellappa, C.I.Wilson and S.Sirohey, "Human and Machine recognition of faces: A survey", Proceedings of IEEE vol.83, vol.5, pp. 705-741,1995.
- A.K.Jain, A.Ross and S.Prabhakar, "An introduction to biometric recognition",IEEE Transaction Circuits System. Video Technologies vol 14, no 1 pp. 4-20,2004.
- Juwei Lu et.al , "Face Recognition Using Kernel Direct Discriminant Analysis Algorithms", IEEE Transcations on Neural Network, vol 14, pp117-126,2003.
- Tai Sing Lee et. al, "Image Representation Using 2D Gabor Wavelets",IEEE Transactions on Pattern Analysis and Machine Intellignece, Vol 18, pp1-13,1996.
- B.Modhaddan and A.Pendland,(1997) "Probabilistic visual learning for object Representation" IEEE Trans, Pattern Ana. Mach, Intelligence,. vol 19, no.7,pp.696-710,1997.
- Linlin Shel et. al, "A review on Gabor Wavelets for face recognition", Pattern Analysis Application, pp 273-292,2006.
-  J.G. Daugman, "High confidence visual recognition of persons by a test of statistical independence", IEEE Trans. Pattern Anal. Mach. Intell. vol.15, no.11, pp 1148-1151, Nov, 1993.
- Hongming Zhang et. al , "Object detection using spatial histogram features", Image and Vision Computing, pp. 1-15,2006.
-  Sanders J.S. and Fabian A.C. (2001), "Adaptive Binning of X-Ray Galaxy Cluster Images", Institute of Astronomy, Cambridge, pp 1-8.
-  Singh S.K. (2003), "A comparative study of various face recognition algorithms", IEEE International Workshop on Computer Architecture for Maching Perception, pp.160-171.
-  Richard A. Davis (2007), "Adaptive Binning: An improved binning method for metabolomics data using the undecimated wavelet transform", Chemometrics and Intelligent Laboratory Systems, pp. 144-154.
Dr.A.Srinivasan, completed his ME, PhD in Computer Science and Engineering at Madras Institute of Technology, Anna University, Chennai. He has finished his Post doctorate at Nan yang Technological University, Singapore. He has 17 years of Teaching and Research Experience in Computer Science and Engineering field and one year of Industrial Experience. At present, he has five Ph.D students working under him. He has published more than 32 Research publications in National and International journals and conferences. He is Editorial board member to Journal of Computer Science and Information Technology [ JCSIT] and a Reviewer to four Reputed International Journals in Computer Science and Engineering field. Currently he is working as Professor in Computer Science and Engineering Department, Sri Sivasubramania Nadar College of Engineering, Anna University, Chennai, India. His field of interests are Digital Image processing and Analysis, Face Recognition and Distributed Systems.
If you are the original writer of this essay and no longer wish to have the essay published on the UK Essays website then please click on the link below to request removal:
More from UK Essays
- Free Essays Index - Return to the FREE Essays Index
- More Technology Essays - More Free Technology Essays (submitted by students)
- Technology Essay Writing Service -find out more about how we can help you
- Example Technology Essays - See examples of Technology Essays (written by our in-house experts)
Need help with your essay?
We offer a bespoke essay writing service and can produce an essay to your exact requirements, written by one of our expert academic writing team. Simply click on the button below to order your essay, you will see an instant price based on your specific needs before the order is processed: