Facial Expression Recognition using Image Processing and Neural Network

Transcription

1 Keerti Keshav Kanchi / International Journal of Comuter Science & Engineering Technology (IJCSET) Facial Exression Recognition using Image Processing and eural etwork Keerti Keshav Kanchi PG Student, Deartment of Electronics and Communication Engineering SD College of Engineering and Technology Dharwad , Karnataka, India -id: keertikanchi@gmail.com Abstract Face recognition is an imortant techniue and has drawn the attention of many researchers due to its varying alications such as security systems, medical systems, entertainment. A face recognition system is one of the biometric information rocessing systems. The develoed algorithm for the facial exression recognition system, which uses the two-dimensional discrete cosine transform(2d- DCT) for image comression and the self organizing ma(so) neural network for recognition urose, simulated in ATLAB. By using 2D-DCT we extract image vectors and these vectors become the inut to neural network classifier, which uses self organizing ma algorithm to recognize familiar faces (trained) and faces with variations in exression. In this aer we have develoed and illustrated a recognition system for human faces using a novel self organizing ma based retrieval system. SO has good feature extracting roerty due to its toological ordering. The facial analytics result for the 25 images of AT &T database reflects that the face recognition rate using one the neural network algorithm SO is 95.05% for 5 ersons. Keywords- 2D-Discrete Cosine Transform (2D-DCT), Facial Exression Recognition, eural etwork, Self Organizing a. I. ITRODUCTIO Face recognition is one of the most successful alications of attern recognition and image analysis which has recently received significant attention, esecially during the ast several years. At least two reasons account for this trend: the first is the wide range of commercial and law enforcement alications, and the second is the availability of feasible technologies after 30 years of research. Face Recognition has gained momentum and ractical vitality in the wake of increased and growing security concerns. A facial recognition and face verification system can be considered as a comuter alication for automatically identifying or verifying a erson in a digital image in as much as the rocessing is carried out on digital still facial images []. The facial exression recognition system was introduced in 978 by Suwa et. Al. The main issue of building a facial exression recognition system is face detection and alignment, image normalization, feature extraction, and classification. There are number of techniues which we use for recognizing the facial exression. This aer resents a novel aroach to recognize face using two dimensional discrete cosine transforms (2D- DCT) and self organize ma (SO) eural etwork as classifier. A block diagram of roosed techniue for facial exression recognition using SO eural etwork is as shown in the fig. In the first stage all the 25 face images are comressed for feature rocessing using two dimensional discrete cosine transform (2D-DCT). When the 2D DCT is alied with a mask, high-coefficients in an image are discarded []. Then the 2D IDCT is alied to regenerate the comressed image, which is blurred due to loss of uality and also smaller in size. In next stage uses a self-organizing ma (SO) with an unsuervised learning techniue which is trained to classify vectors into grous to recognize if the subject in the inut image is resent or not resent in the image database. After training all the 25 face images, we now take a single inut image, comress it using 2D- DCT and regenerate it using IDCT. All the 25 trained images and untrained inut image are simulated. The untrained inut image is comared with all trained images, if the face image is classified as resent, the best match image is found in the training database using minimum absolute deviation and that image is dislayed otherwise if the image is not found then image is not found in the database is dislayed. ISS : Vol. 4 o. 05 ay

2 Keerti Keshav Kanchi / International Journal of Comuter Science & Engineering Technology (IJCSET) Fig. Block Diagram of Facial Exression Recognition using SO eural etwork II. DISCRETE COSIE TRASFOR For feature extraction 2-D Discrete Cosine Transform (2D-DCT) is used. A discrete cosine transform (DCT) exresses a seuence of finitely many data oints in terms of a sum of cosine functions oscillating at different freuencies. The use of cosine rather than sine functions is critical in these alications: for comression, it turns out that cosine functions are much more efficient, whereas for differential euations the cosines exress articular choice of boundary conditions [2]. The DCT, and in articular the DCT-II, is often used in signal and image rocessing, esecially for lossy data comression, because it has a strong "energy comaction". ost of the signal information tends to be concentrated in a few low-freuency comonents of the DCT.DCT is real valued and rovides a better aroximation of a signal with fewer coefficients. The 2D-DCT of an matrix A is defined as follows: 2 2 m n B A cos cos,0,0 (2.) mn m0 n0 2 2 The values B are the DCT coefficients. The DCT is an invertible transform, and the 2D-IDCT (2D Inverse- DCT) is defined as follows: A mn 0 0 B cos 2m 2n 2 cos 2,0 m,0 n (2.2) The values and in euation (2.) and (2.2) are given by: 2,, 0 2 The roosed techniue uses the DCT transform matrix in the ATLAB Image Processing Toolbox. This techniue is efficient for small suare inuts such as image blocks of 8 8 ixels. The transform matrix T is given by:,, 0 (2.3) ISS : Vol. 4 o. 05 ay

3 Keerti Keshav Kanchi / International Journal of Comuter Science & Engineering Technology (IJCSET) T 2, cos 0, 0 2 2, The DCT transformation matrix is imlemented in ATLAB and the feature is extracted. The 2D DCT reduces the size of the data significantly by transforming the image from the satial reresentation into the freuency domain. The lower freuencies are characterized by relatively larger magnitude while the higher freuencies have smaller magnitudes [3]. The higher freuency comonents are ignored as it does not significantly affect the accuracy of work. In short, the 2D DCT coefficient of lower freuency values catures the most relevant information of the facial exressions. A. 2D-DCT Image Comression The roosed design techniue calculates the 2D-DCT of the image blocks of size 8 8 ixels using 8 out of the 64 DCT coefficients for masking. The other 56 remaining coefficients are discarded (set to zero). The image is then reconstructed by comuting the 2D-IDCT of each block using the DCT transform matrix comutation method. Finally, the outut is a set of arrays. Each array is of size 8 8 ixels and reresents a single image [4]. Emirically, the uer left corner of each 2D-DCT matrix contains the most imortant values, because they corresond to low-freuency comonents within the rocessed image block. III. SELF ORGAIZIG AP Self organizing ma also known as kohonen ma is well known artificial neural network. It is unsuervised learning rocess in which the learning is based uon the inut data which is known as unlabeled data and is indeendent of the desired outut data. Self organizing ma can also be termed as a toology reserving ma [5]. There is a cometition among the neurons to be activated and only one neuron that wins the cometition is fired and is called the winner. Kohonen rule is used to learn the winner neuron and neurons within a certain neighborhood of the winning neuron. This rule allows the weight of neuron to learn an inut vector so this makes it erfect for recognition. Hence in this system SO is used as classifier. The SO network used in this system contains nodes ordered in two dimensional lattice architecture and each node has 2 or 3 neighbor nodes. SO has three hases of life cycle: learning hase, training hase and testing hase. A. Unsuervised Learning During the learning the neurons having weight closest with the inut vector declare as winner [6]. Based on winning neuron weights of all neighborhood neurons are adjusted by an amount inversely roortional to the Euclidean distance. The learning algorithm is summarized as follows:. Initialization: Choose random values for the initial weight vectors w j 0, the weight vector being different for j =, 2,...,l where l is the total number of neurons. T n w i w i, w i 2,..., w il (3.) 2. Samling: Draw a samle x from the inut sace with a certain robability. Where x T n x x, 2, x l (3.2) i x at time t, 0 < t n by using the minimum distance Euclidean criterion: i x arg min x( n) w, j,2... l (3.3), 0 3. Similarity atching: Find the best matching (winning) neuron j j 4. Udating: Adjust the synatic weight vector of all neurons by using the udate formula: wj ( n ) wj ( n) ( n) hj, i( x) ( n) x( n) wj ( n) (3.4) n is learning rate arameter, and h i xn nand are xn winning neuron. Both h i (2.4) j, is the neighbourhood function centred around the j, varied dynamically during learning for best results. 5. Continue with ste 2 until no noticeable changes in the feature ma are observed. B. Training During the training hase feature vector are resented to the SO one at a time. For each node net determines the outut unit that is best matches for current inut samle. The weight vector for the winner is adjusted with resect to the learning algorithm described in learning hase [7]. At the end of this hase each ISS : Vol. 4 o. 05 ay

4 Keerti Keshav Kanchi / International Journal of Comuter Science & Engineering Technology (IJCSET) node has two values: total number of winning times for the subject resent in the database, and total number of winning times for the subject not resent in the database. C. Testing During the testing hase firstly every inut vector is comared with the all the SO nodes and then the best match is found based on minimum Euclidean distance as given in euation (3.3). After that final outut result is shown. IV. EXPERIETAL WORK A. Image Database The training database consists of 25 images, containing five subjects and each subject having 5 individual face images with 5 different facial exressions. The image databases for of 25 subjects with different facial exressions are as shown below: (a) (b) Fig. 2 Training and testing image database (a) Image database for training of 25 subjects with 5 different facial exressions. (b) Untrained image for testing. B. Validation Techniue All the face images were resized to 8 x 8 ixels and saved; the next ste was to comress them by alying the 2D blocked DCT. When the 2D DCT is alied with a mask, high-coefficients in the image are discarded. Then the 2D IDCT is alied to regenerate the comressed image, which is blurred due to loss of uality and also smaller in size. For the image data to be inut into the neural network, it should follow the form of only one column, desite the number of rows. Currently, all the resized and DCT comressed face images are in the form of 8 x 8 ixels. Hence the image data needed to be reshaed from an 8 x 8 matrix to a 64 x array for it to be used both for the inut and training database of the neural network. SO s were found to be efficient for image data management and roved to be an accurate closest matching techniue of untrained inut images with trained database of images. For the design of SO, a set of 25 image data, 5 different subjects with 5 different facial exressions for the training database was loaded into ATLAB [8]. A SO was then created and the Parameters for the SO network were selected to be a minimum and maximum oint for each row on vector; training database. There were 64 minimum and 64 maximum oints selected altogether. After the SO neural network was created, it was trained for 2000 eochs. The figure 3 is the SO layer weight for the 25 face images in the training database and figure 4 is the SO weight vectors is as shown below: ISS : Vol. 4 o. 05 ay

5 Keerti Keshav Kanchi / International Journal of Comuter Science & Engineering Technology (IJCSET) Fig. 3 SO layer weights for the 25 face images in the training database. Fig.4 SO weight vectors After the SO neural network was trained and simulated for the 25 images in the training database, the SO neural network was then simulated for the single inut face image. After the SO neural network is simulated for the inut face image, the image in the training database which is the closest match by the SO neural network for the inut face image is found by finding the minimum absolute deviation. After the closest matched training database images are found, they are then classified. Classification of the subject is the answer of the face recognition system. C. Determining Otimal umber of Eochs for Training Eochs are neural network training arameters. They are defined as one comlete cycle through the neural network for all cases, which resent the entire training set to the neural network. Each time a record goes through the net, it is one trial, one swee of all records is termed as an Eoch. Less number of eochs used for training leads to less training time for the training data set. The goal is to find the otimal number of eochs for training which will roduce accurate neural network results and at the same reuire the least amount of time for rogram execution. The SO neural network was tested to determine the otimal number of eochs to be used for neural network training. This test was erformed by varying the number of eochs and network training time to find the best ossible recognition rate. Table I shows the tabulated results for the test. The training database for this test consists of 25 face images of 5 subjects, each subject having 5 individual images with different facial exressions. ISS : Vol. 4 o. 05 ay

6 Keerti Keshav Kanchi / International Journal of Comuter Science & Engineering Technology (IJCSET) Table I. Comarison of number of eochs vs. network training time and Recognition rate umber of eochs etwork Training Time (in sec) Recognition Rate (%) Table I shows the tabulated results generated by varying the number of eochs. The most efficient number of eochs for training and with resect to fastest training time is 2000, with achieves best ossible recognition rate as 95.05%. V. COCLUSIO This aer has resented a novel facial exression recognition techniue that uses features derived from DCT coefficients, along with a self organizing ma (SO)-based classifier. The 2D-DCT and SO neural network are the heart for the design and imlementation of efficient facial exression recognition system. The system was evaluated in ATLAB using an image database of 25 face images, containing five subjects and each subject having 5 images with different facial exressions. After training for aroximately 2000 eochs the system achieved a recognition rate of 95.05% for fastest network training time. The system having less comutational reuirement this make system well suited for low cost, real-time hardware imlementation. VI. ACKOWLEDGET It is a great leasure to exress sincere and humble gratitude to the guide Dr. Vijaya C for valuable guidance and encouragement given during the course of work. The author would also like to extend sincere thanks to the Princial, HOD and to all teaching and non-teaching staff of Electronics and Communication Engineering deartment of SD College of Engineering and Technology, Dharwad, Karnataka, India for their constant suort. REFERECES [] Jawad agi, Syed Khaleel Ahmed and Farrukh agi, A ATLAB based Face Recognition System using Image Processing and eural etworks, 4th International Collouium on Signal Processing and its Alications, arch 7-9, 2008, Kuala Lumur, alaysia. [2] Dinesh Chandra Jain and V. P. Pawar, A ovel Aroach For Recognition Of Human Face Automatically Using eural etwork ethod, International Journal of Advanced Research in Comuter Science and Software Engineering, Volume 2, Issue, January 202. [3].andini, P.Bhargavi and G.Raja Sekhar, Face Recognition Using eural etworks, International Journal of Scientific and Research Publications, Volume 3, Issue 3, arch 203. [4] Digital Image Processing - R.C. Gonzalez Richards and Xiuing Jia, enlarged edition, Sringer-Verlag, 999. Woods, Addison Wesley, 992. [5] Simon Haykin, eural etwork-a Comrehensive Foundation, 2 nd Ed., Pearson Education, 2004 [6] Steve Lawrence, C. Lee Giles, Ah Chung Tsoi and Andrew D. Back, Face Recognition: A Convolutional eural etwork Aroach, Acceted for ublication, IEEE Transactions on eural etworks, Secial Issue on eural etworks and Pattern Recognition. [7] Angel oe artinez-gonzalez and Victor Ayala-Ramirez, Real Time Face Detection Using eural etworks, 20 0th exican International Conference on Artificial Intelligence. [8] Abdallah S. Abdallah, A. Lynn Abbott, and ohamad Abou El-asr, A ew Face Detection Techniue using 2D DCT and Self Organizing Feature a, Proceedings Of World Academy Of Science, Engineering And Technology Volume 2 ay ISS : Vol. 4 o. 05 ay