# Introduction Face recognition [1][2] [3] is playing a key role in the research areas of pattern recognition and computer vision due to its capable applications such as personal identification, bankcard verification, automated surveillance etc. So far, various face recognition systems have been proposed but still research is going on in the field of face recognition as it has various inherent complexity associated with environmental changes and subject actions (facial expressions and head pose variations), etc. The performance of face recognition systems mainly varies corresponding to the environment where face images are taken. These Face recognition approaches normally use visible spectrum images for recognition because they provide clear representation of facial features and face texture to differentiate between two individuals. With the growth of internet and multimedia, compression techniques have become the popular research area in the field of Image processing. But, there are very few researchers have concentrated on to investigate the effect of compression on face recognition, even though the images are mainly stored and/or transported in a compressed format. Image Compression is an area of face recognition which needs more attention since the adoption of face recognition as part of the e-passports which use face images as one of the three identifiers along with fingerprints and iris scans being the other two. Mainly, the compression of images are essential in face recognition due to the advantages such as ability to store compressed images on smart cards and low-capacity chips, easy to transmit image to a distant server for verification and improved computational speed [4]. Moreover, image acquisition equipment also often delivers the compressed images as its output. So, to work directly in existing compressed domain has its valid reasons as well. For the stored (or derived) image of the unknown individual to be identified it has to be compared to one or more images of known individuals. Decompressing all those images (including the unknown one) is very computationally intensive. Thus, avoiding full decompression process could be very advantageous when compared to the existing techniques [5]. # II. # Background # Image Compression Technique Using Wavelets The most widely used standard image compression techniques are JPEG [6] and JPEG2000 [7]. These image compression standards have been employed in face recognition systems in past few years. Usually, compression is vital for any reasonable implementation where a large quantity of images need to be stored. All modern image coders are considered to be transform coders. The Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT) are examples of those transforms methods. Both JPEG and JPEG2000 have been using the general transform coding scheme as shown in Figure 1. The input signal is first encoded as bit stream using an invertible transform, and then quantized and entropy coding has been done. The rate distortion unit controls the quantization to minimize the distortion within the available bit rate. These stages have an impact on the image quality achieved in terms of compression ratio. A better transform will provide improved compression performance [8]. In decompression process, the data must be entropy decoded first, inverse quantized and then inverse transformed in order to reconstruct an original image. This resultant image may be with a poor quality due to information discarded during compression. In figure 1, point D represents image pixels and point A represents fully compressed domain. The resulting reconstructed images will have a checker-board or the smear effect. As the transform coefficients that represent higher frequencies are more and more discarded (or are rounded to lower precision) with higher compression rates, the images become more and more low-pass filtered. This is clearly observed in the JPEG2000. For example at 0.2 bpp, the finer details of the face (like wrinkles) were eliminated in the reconstructed image. Thus it is observed that, low-pass filtering can greatly affect the recognition results [9]. The magnitude of compression is mainly calculated by compression ratio, expressed in the form of bits per pixel (bpp). The performance of the compression technique can be effectively measured by Peak Signal to Noise Ratio (PSNR), and calculated as: [dB] where n is the number of bits per pixel in the original image and RMS is the Root Mean Square Error defined as: where Ii is pixel value in the original image, I'i is corresponding pixel value in the reconstructed image and N is the total number of pixels in the image. The basic image compression technique usually takes high computational time for stored image of the unknown individual and it can be recognized if it is compared to one or more images of known individuals. Moreover, decompressing all those images will be very difficult; this leads to a novel idea of image compression technique which will avoid the difficult position in the decompression domain. # III. # Related Work In Ekenel, H.K. et al. [10], the face recognition system that can be deployed for real-life smart interaction applications have been described. This system is capable of developing a fast and robust fully automatic face recognition system. The algorithm is mainly based on the appearances of local facial regions that are represented with the coefficients of discrete cosine transform. There are three fully automatic face recognition systems have been formulated based on this algorithm. The first device is the "Door monitoring system" which is formulated based on this algorithm that records everything which is entering into the room and identifies the subjects. The second device is "portable face recognition system" which recognizes the user of a machine. The third system "3D face recognition system" which operates fully automatic face recognition on 3D range data. Rabab M. Ramadan et al. [11] have presented Face recognition comprises of many applications which are ranging from security access to video indexing by content. Here they described about the VQ-based automatic face recognition system and which analyze the feature selection effects, codebook size and feature dimensionality on recognition performance in the VQbased framework. In specific, the DCT-based feature vectors are examined in such a system. Directly the recognition can be performed on the bit stream of the DCT based compressed images. The system comprises of three parts: a preprocessing step which is accomplished to segment the face, then the feature selection process and finally the classification. A database which comprises of 500 images could be recognized with high accuracy. In a study conducted by Delac in 2007 [12], the feasibility of implementing face recognition algorithms directly into JPEG2000 compressed domain was examined and CDF 9/7 wavelet was used in the study. The Standard face recognition algorithms PCA, ICA and LDA were fed with CDF 9/7 wavelet and also FERET database was used. It was proved that no significant drops were observed and the authors claimed that face recognition algorithms can be directly implemented in JPEG2000 compressed domain. The authors also recommended further research for finding a way to intelligently extract feature vectors within compression scheme or finding another mechanism for image matching in order to improve recognition performance. This proposed method continues the study reported in this paper. IV. # PROPOSED METHODOLOGY The proposed methodology uses the entropy points as the input to the face recognition system as in PCA. The original image is transformed using the DWT and quantization and entropy coding is done on the coefficients. The proposed approach eliminates major part of the decompression phase. The clear approach to transport face recognition into the compressed domain is to stop the decompression process before the inverse transformation and to utilize coefficients as input to face recognition approaches. The inverse quantization as well as the inverse transformation is removed. There is no change in the performance when using the inversely transformed coefficients. When the inverse quantization is eliminated some additional computation time was saved. For the experimental observation, all the coefficients available after the entropy decoding were used. In case of the normalized 128 x 128 sized images, all 1195 coefficients per image were used as input to PCA [13]. # Improved Matching by Kd tree The K-dimensional tree (kd-tree) is a spacepartitioning data structure for organizing points in a kdimensional space. This method is a widely used data structure in various applications like exploration which involves multidimensional search key (e.g. range searches and nearest neighbor searches). The kd-tree methods are considered as a special case of BSP trees. In this proposed approach, kd tree is used for improved matching of the face recognition algorithms. Kd-tree Algorithm: Given a list of n points, the following algorithm will construct a balanced kd-tree containing those points. Algorithm Build Kd tree (P, depth) Input: A set of points P and the current depth. Output: The root of a kd-tree storing P. 1. If P contains only one point 2. then return a leaf storing this point 3. Else if depth is even 4. then Spilt P into two subsets with a vertical line l through the median x coordinate of the points in P. Let P 1 be the set of points to the left of l or on l , and let P 2 be the set of points to the right of l. # 5. Else Spilt P into two subsets with a horizontal line l through the median ycoordinate of the points in P. Let P 1 be the set of points below l or on l , and let P 2 be the set of points above l. 6. V left -BUILDKDTREE(P 1, depth+1) 7. V rightt -BUILDKDTREE(P 2, depth+1) 8. Create a node V storing l , making V left the left child of V , and make V rightt the right child of V . ? Querying an axis-parallel range in a balanced kdtree takes O (n 1-1/k +m) time, where m is the number of the reported points, and k the dimension of the kd-tree. Thus from the complexity information, it is found that compared to Euclidean distance which was used in earlier studies for recognition purpose, the proposed kd-tree matching technique takes less time. V. # Experimental Setups And Results # Database and protocol To implement the proposed idea, a Standard FERET data [13] set is utilized in this paper for evaluating the proposed technique which includes the data partitions for recognition tests, as represented in [14]. This dataset contains 1,196 images and also four sets of query images that are matched to the dataset images in identification phase. The fb probe set consists of 1,195 images of subjects obtained at the same time as dataset images with the only dissimilarity being that the subjects were told to suppose a various facial expression. The fc probe set consists of 194 images of subjects with various illumination cases. The dup1 (duplicate I) set consists of 722 images taken somewhere between one minute and 1,031 days after the dataset image was obtained, and dup2 (duplicate II) set is nothing but a subset of dup1 which consists of 234 images obtained at minimum18 months after the dataset image was obtained. Every image in the data # Preprocessing There are several well known appearancebased subspace face recognition algorithms like Principal Component Analysis (PCA) [15], Linear Discriminant Analysis (LDA) [16] and Independent Component Analysis (ICA) [17]. The Original FERET images are initially transformed spatially to obtain the eyes at the fixed points depending on a ground truth file of the eye coordinates provided with the original FERET data. Each and every image is then cropped to size of 128 × 128 pixels and an elliptical mask was utilized to mostly to remove the background. At last, image pixel values were histogram equalized to the sort of values between 0 and 255. These preprocessing processes are performed on each image earlier to carry out the experiments even before compression. In training the PCA algorithm, a subset of classes that there are three images per class was utilized. It is identified that there are 255 such classes which represent 255 various persons, thus the training set contain total of 675 images with 3 images per person (M = 675, c = 225). The outcome of this process is that this percentage of overlie has on the performance of this algorithm requires additional examination. The result of PCA algorithm according to the theory is (M -1 = 674) meaningful eigenvectors. The FERET suggestion is assumed and reserved the top 40% of those, consequence in 270-dimensional PCA subspace W (40% of 674 = 270). It was obtained that 97.85% of energy was preserved in those 270 eigenvectors. The performance of the proposed approach compared to the existing image compression algorithms were evaluated by the performance measure called the Normalized Recognition Rate (NRR). In Figure 3, it shows the graphical representation of the NRR of the proposed method compared to the existing method. It is clearly observed from the graph that the proposed approach shows the improvement in the NRR compared to the existing recognition technique. The computational time of the proposed approach using the kd tree method is compared with the existing techniques in figure 4. The kd-tree takes very less computation time when compared to the recognition rate of the existing approach. # CONCLUSION This paper presents systematic approach of implementing face recognition methods directly in compressed domain. All the aspects that can occur in such circumstances, ranging from the effects of compression on recognition rate to technical implementation issues like computational time and storage requirements savings were analyzed. The experimental results proved that implementing the standard recognition method such as PCA directly in compressed domain will not deteriorate recognition results. All the experimentation results in compressed domain illustrated in this paper were compared to the results in pixel domain i.e. using either uncompressed or fully decompressed images. Moreover, the proposed approach uses the kdtree method for the face recognition comparison. This algorithm effectively reduces the computational time to a very great extent. Thus the proposed approach provides The scope for further improvement of this work will consist of analyzing the influence of proposed methodology on the result of performance for known face recognition algorithms using Linear Discriminant Analysis (LDA) and Independent Component Analysis (ICA). This could be implemented by testing on various images from different databases for generalization. 1![Figure 1: Block diagram of compression /decompression in transform coding](image-2.png "Figure 1 :") 3![Figure 3: Face Recognition in Compressed Domain It is very vital to define the compressed domain as it plays a key role in this research. The Compressed domain is any point in the compression / decompression procedure after transform and before inverse transform.The clear approach to transport face recognition into the compressed domain is to stop the decompression process before the inverse transformation and to utilize coefficients as input to face recognition approaches. The inverse quantization as well as the inverse transformation is removed.There is no change in the performance when using the inversely transformed coefficients. When the inverse quantization is eliminated some additional computation time was saved. For the experimental observation, all the coefficients available after the entropy decoding were used. In case of the normalized 128 x 128 sized images, all 1195 coefficients per image were used as input to PCA[13].](image-3.png "Figure 3 :") ![Constructing a static kd-tree from n points takes O (n log 2 n) time if an O (n log n) sort is used to calculate the median at each level. The complexity is O (n log n) if a linear median-finding approach. ? Inserting a new point into a balanced kd-tree takes O (log n) time. ? Removing a point from a balanced kd-tree takes O (log n) time.](image-4.png "?") ![size 384 × 256 pixels and represented in grayscale.](image-5.png "") 3![Figure 3: NRR evaluation of the proposed approach](image-6.png "Figure 3 :") 4![Figure 4: Comparison of the Computational Time VI.](image-7.png "Figure 4 :") MayFace Recognition in compressed domain based on wavelet transform and kd-tree matching ©2011 Global Journals Inc. (US) MayFace Recognition in compressed domain based on wavelet transform and kd-tree matching ©2011 Global Journals Inc. (US) MayFace Recognition in compressed domain based on wavelet transform and kd-tree matching ©2011 Global Journals Inc. (US) MayFace Recognition in compressed domain based on wavelet transform and kd-tree matching ©2011 Global Journals Inc. (US) MayThis page is intentionally left blankFace Recognition in compressed domain based on wavelet transform and kd-tree matching ©2011 Global Journals Inc. (US) * Face Recognition: A Literature Survey WZhao RChellappa ARosenfeld PJPhillips ACM Computing Surveys 35 4 2003 * A Survey of Biometric Recognition Methods KDelac MGrgic Proc. of the 46th International Symposium Electronics in Marine, ELMAR-2004 of the 46th International Symposium Electronics in Marine, ELMAR-2004Zadar, Croatia June 2004 * Handbook of Face Recognition S.Z. Li, A.K. Jain 2005 Springer New York, USA * Effects of Compression and Individual Variability on Face Recognition Performance DPMcgarry CMArndt SAMccabe DPD'amato Proc. of SPIE of SPIE 2004 5404 * Towards Facre Recognition in JPEG2000 Compressed Domain KDelac MGrig Grgic Proc. Of the 14th International Workshop on Systems, Signals and Image Processing(IWSSIP) Of the 14th International Workshop on Systems, Signals and Image essing(IWSSIP) June 2007 * The JPEG Still Picture Compression Standard GKWallace Communications of the ACM 34 April 1991 * The JPEG 2000 Still Image Compression Standard ASkodras CChristopoulos TEbrahimi IEEE Signal Processing Magazine 18 5 September 2001 * Effective Image Compression using Evolved Wavelets UniGrasemann RistoMiikkulainen ACM 1-59593-010- 8/05/0006 GECCO'05 June 2005 * Face Recognition in JPEG and JPEG2000 compressed domain KDelac MGrgic Grgic Image and Vision Computing 27 2009 * Face Recognition for Smart Interactions HKEkenel JStallkamp HGao MFischer RStiefelhagen IEEE International Conference on Multimedia and Expo 2007 * Face Recognition Using Particle Swarm Optimization-Based Selected Features MRabab RehabFRamadan Abdel -Kader IEEE International Symposium on Signal Processing and Information Technology (ISSPIT) 2009 585 * Fusing face recognition from multiple cameras JChangboHarguess JKHu Aggarwal Dept Workshop on Applications of Computer Vision (WACV) Austin, TX, USA 2009 of ECE, Univ. of Texas at Austin * The FERET Database and Evaluation Procedure for Face Recognition Algorithms PJPhillips HWechsler JHuang PJRauss Image and Vision Computing 16 5 1998 * DMBlackburn JMBone PJPhillips FRVT 2000 Evaluation Report 2001 * Eigenfaces for Recognition MTurk APentland Journal of Cognitive Neurosicence 3 1991 * Eigenfaces vs. Fisherfaces: Recognition using Class Specific Linear Projection PNBelhumeur JPHespanha DJKriegman Proc. of the 4th European Conference on Computer Vision, ECCV'96 of the 4th European Conference on Computer Vision, ECCV'96Cambridge, UK April 1996 * Face Recognition by Independent Component Analysis MSBartlett JRMovellan TJSejnowski IEEE Trans. on Neural Networks 13 6 2002