Face Detection This projects goal was to implement a neural network based face detector as outlined in this paper:

Rowley, H., Baluja, S., and Kanade, T., "Neural Network-Based Face Detection", Proc. IEEE Conf. on Computer Vision and Pattern Recognition, San Francisco, CA, pp. 203-207. 1996. There is a newer version of the paper with better figures here < http://www.vision.caltech.edu/CNS179/papers/Kanade98.pdf >. The authors of the paper have set up a cool web demo < http://demo.pittpatt.com/ > that allows anyone to upload a picture and try out their detector. Abstract The main aim of the project is to design and develop an application that provides data security with the help of face detection system. The reason behind developing this project is to protect the information of a individual person. Face recognition systems are computer based security systems that automatically identifies human faces. This system uses various detection algorithms like Forward lifting and Discrete wavelet transform for face detection. Recently, Face detection technology is capturing much attention of the network multimedia society access. Network access control through face detection technology eliminates various security problems and increases user friendliness in human computer interaction. This report presents the general frame work for the face detection system. HAAR algorithm is a face detection algorithm which has been developing in the recent years. This algorithm is mainly utilized to detect the face of a person. It mainly involves in the computer and physical access control, various security applications, criminal investigation and real-time subject identification and authentication. Face Detection algorithm can be categorized in to two functional modules such as a face image detector and face recognizer. Both the face recognizer and face detector uses feature extractor which transforms the facial image pixels to useful vector representation. Pattern recognizer searches the image in the database to find the best match to the incoming to the face image. A wavelet that is having a different example in the functional and mathematical analysis is referred as, Discrete Wavelet Transform (DWT). This wavelet will have the capacity to detain the data of the position and the occurrence of the wavelet. This process will be defined as the Temporal resolution. Comparing to the other transforms of wavelet like Fourier, it is the major advantage of this Discrete Wavelength Transform. Alfred Haar, who is mathematician from Hungarian, has invented the initial discrete wavelet transform. Overview This project seeks to detect upright human faces in grayscale images with the help of a neural network. The retinally connected neural network (NN) examines small windows in the image and outputs a confidence factor corresponding to whether or not it thinks the window contains a face. Heuristics are then used to clean up the output of the network. This algorithm is not real-time as it can take up to an hour to analyze even a moderately sized an image. I created my NN detector in Matlab using its NNtoolbox. I followed the paper pretty closely; however my implementation only used 1 NN with only 1 layer of hidden nodes. The paper used

multiple networks with multiple hidden layers. My network achieved a face detection rate of 41.3% and a false positive rate of 0.00044%. These percentages are much lower than the paper's, therefore, there are some parameters of the algorithm that can still be optimized. Introduction In todays environment there is a need to maintain the security or information is complex task. The criminals are taking benefit of some fundamental flaws in existing system. This systems dont verify for who we are instead it verifies what we have like keys, Pin number, password etc. Hacker can easily acquire this information and can access ones account. To prevent this problem Biometrics technology came in to existence. It is mainly based on field namely biometrics. Biometric techniques are the methods of verifying or identification of identity of living person based on the various physiological features like finger prints or facial characteristics. As this method detects the person by its physical or biological features, they are difficult to forge. Among them face detection is one of the method that have various advantages of both accuracy and low intrusiveness. Several algorithms had been proposed for face detection system. [Bruno et al, 2003] The Face detection encounters multifarious variations. So various detection algorithms are developed to cope up with such problems. Among them Eigen face and neural networks are familiar approaches. The Eigen face implements Karhonen-Loeve (KL) transform for feature extraction. It decreases dimension of facial features and sustains discriminative perceptive power. In neural network, some variants of algorithm operate on feature extraction. It mainly presents effective scheme likelihood densities the patter identification phase. Patter recognizer classify the incoming feature vector two image classes that is face image and non face images. Usually the face detector produces only an approximate location of the face due to various kind of variations. All the functional modules of the system like face detector and face recognizer implements the schematics of feature extractor and pattern recognizer. [Qingqing, 2009] Face detection attains to find a person face in the image. It can be accomplished by using various methods like movements, blurred human shapes and skin tones. Initially it locates the heads and finally it locates the eyes. Matrix will be developed based on various features of the human face. The process of defining the matrix varies as per algorithm. This matrix will be compared t matrices in the data base and generates similarity score. Based on this it shows the results. Eigen face increases the efficiency of feature extraction. Covariance matrix of Eigen vectors forms the KL bases. The few Eigen values of high dimensional space have large values. The popular forward lifting algorithm is used to identify face images directly. Forward lifting algorithm is mainly applied to pattern recognition than feature extraction phase to lessen the complexity. Dimensional, horizontal, vertical and absolute of the facial image is determined by forward lifting algorithm before it fed in the face detector. [Jong-Hwan Kim et al, 2009] Back ground Survey The main purpose of face detection systems is to provide security in the fields it is being implemented. This serves as identification of exact individual that is preloaded in the system of the organization where it has been implemented. It facilitates in observing the usage of the system by the authorized person and if it is unauthorized person access is denied. It is mainly operated in boarder security problem and surveillance areas where a small Camera can catch the high quality image of the person in the field [ Ana, 2001]. Face detection system provides a way to law to central bureau of investigation, British Secret agency and government agencies to maintain the records of people, who involved in crimes. It helps in easy identification of the culprit in the crime. Face detection is a unique system that facilitates the crime department to identify the individual person in the investigation where finger prints is not allowed. There are different biometric techniques in the face detection systems, from which two biometric systems are combined, and then the accuracy of the results that are

generated is high in crime investigation. Hence in most of the investigation cases, the department uses a combination of biometric techniques that reduces the time of identifying the actual offender. A small camera and transmission wave helps to identify the person in the field with this system. Hence the police officers that are equipped with PDAs system can easily identify the facial features of the individual and search the results of that individual in the law enforcement which reduces the risk and time [ Eric, 2010]. The face recognition technique is also called as Face detection. The face detection system will not consider the age, movements and expressions of the person. The detection starts with the inner features of the face. The face detection techniques will mostly calculates the skin tone and features of the face like the structure of the bones, figure of the head and also includes the moment of the persons in the detecting clippings. Face detection will completely work as a tool in the system of face recognition. The process of face detection will be carried out with two windows. One windows is used to execute the scanned photographs and the second window is used to display the matching images. The algorithm of the face detection characterizes into regions of the images that contain the faces using the bounding boxes. In general, there are two face detection techniques. First one is learning based technique and the second one is appearance based technique. Most of the algorithms used to detect the face depending on the appearance of the person face[Jain et al, 2005]. In early days, many areas like government sectors, military sectors and the police department experienced a tough stage with the security issue of identifying the person. These problems have got a solution with the help of the face detection technique. The general face detection techniques are derived into two matching techniques. The first one is the geometric feature matching technique and this technique used to access with the facial characteristics of the person. The initial person eyes, nose and ears will be viewed by matching with the related persons. The features of the derived which are parallel with the source person will be produced as the final result. As the parts of the body are considered for the matching process, this technique is also called as Segmentation technique. The second method is the template matching technique. Here the body parts will not be considered. The entire face will be derived as a bi-dimensional array. This array of the face will be compared with the related persons. Hence it is also referred as the Non-segmentation technique [Amira et al, 2005]. The various biometric techniques categorized into Hand, face, Finger prints, irises, physical and behavioral categories. Based upon the categories the biometric techniques have been derived. The initial biometric technique used is Finger prints of a person. This technique comes under the hand category. This technique is first introduced by the people of Egypt. They used clay seals for identifying the finger prints of a person. In most cases, this biometrics is officially used in exports and imports of the goods. The thumb prints will be given on the source to view by the destination. Later on, it is also implemented by Chinese people. Here, this biometric technology is utilized to identify the foot prints. The first experiment has been examined on the children foot print. It is the second biometric technique next to finger print. In the next stage, this biometric is used to scan the Iris. The Iris is used to observe the circular ring of the eye. This is a most typical technology used to identify the person. Another biometric technique is the scan of Retina. It is used to examine the inner blood cells of the person. Face detection comes to the category of face biometrics. It is used to detect the person by considering his face structural features. The next category is behavior category. In this, the Signature authentication can be observed. Mostly, this technique is used in the banking sectors, to identify the authorized signature of the customer. The next technique is used to recognize the Voice of the person. In the same way, there are many other biometric techniques. Among these all techniques, the face detection biometric technique is becoming a major advantage in many security sectors. The detailed explanation of this technique is given below [Chris, 2005].

Generally, a person can be recognized by observing his facial features. This is the main theme for inventing the face detection technique. Face recognition system can also be referred as the Facial detection technology. The systems which are used for the face detection are using various high qualified technologies. The techniques like scan the images of the individuals and matching with the related persons need high software programs and high resolution cameras. Hence, it is a highly expensive process. The face detection technology is the most popular techniques among all the biometrics methodologies. In the previous days, this trend has worked on the black and white images and also, it can able to identified with unmoving images. But now, this technology is used to identify the persons of color image and videos. The moving and unmoving images are able to rectify by this technology. It is making the persons to make the perfect decisions by observing the detector. The face detector is using in many application areas to identify the fraud activities and criminal investigation. The latest face detectors are having the capacity of scanning the persons from longest distances. Hence, these are mostly using in the higher organizations. There are some disadvantages with other biometrics like finger prints and voice recognition. There is no assurance that all the persons approaching a bank will give the original finger prints. Also, when there is a frequency and wavelength detection to identify the voice, it cannot estimate the exact result. To overcome these problems, the face detection technology is being used [Kelly, 2011]. Experiment 1. Gather a database of face images. I got mine from the Yale < http://cvc.yale.edu/projects/yalefaces/yalefaces.html > and CalTech < http://www.vision.caltech.edu/htmlfiles/archive.html > databases. The faces need to be upright and be 20x20 pixels in size (with approx 12 pixels between the centers of the eyes). The point halfway between the center of the eyes and the upper lip should be in the center of the image. The paper then created 15 images from each face by mirroring, rotating, scaling, and translating. I only mirrored each face. I ended up with 624 face images. 2. Pre-process the faces to eliminate some lighting variation. This involves masking the corners, fitting a Note that the 4 upper hidden nodes each take only a 10x10 quadrant of the input window as their input. The middle 16 hidden nodes each take only a 5x5 pixel region. And the bottom nodes take a 5x20 pixel horizontal slice of the window. Because Matlab does not allow the inputs of a NN to be different, I had to do an ugly and expensive hack. The entire image was fed to all inputs, then the unwanted parts of the image were masked out of the input weights. This masking was done after each iteration of training. This probably considerably slowed the network down. I was essentially carrying around 9000 network weights for no reason. (Only 1527 weights mattered). 5. Run the network on images that are known to contain no faces. I used CalTechs background database < http://www.vision.caltech.edu/html-files/archive.html > for this. Note that the network must be run on all pixel locations in an image pyramid where the images are scaled by a factor of 1.2. The query window must also be

put through the same pre-processing as the training images. Save the false positives to add to the non-faces training set. Go to step 4. This process of retraining with false positives is called bootstrapping. Conclusion In 1888, the first algorithm has been discovered to make the perfect function of face detection system. After that many algorithms have been discovered and implemented. But no algorithm has able to make the perfect function of the system for identifying the person. There are two newly introduced techniques used to implement the system without having any difficulties. They are Elastic graph matching and Eigen face methods. But there are some trouble shots with Eigen face technique. When the image is being evaluated it is making difficulty with the pixel of the images. In this process, there should be a good alignment of the photograph. This is a man kind effort. But, the result after the alignment will be appropriate and adequate. The database will store when a new image will be given in the system. In the Elastic graph method, the image which is inserted for the evaluation will be automatically stored in the database. Here, there is an advantage of Gabor wavelet. So, the pixel of the images will not be considered in the elastic matching technique. It is a time consuming process. It has the complexity with the results evaluated from this method. Thus in many situations, the Eigen face technique will be used than the Elastic face matching technique. There is another wavelet transform called 2-D Gabor wavelet transform. In this the images will be rectified with the graph format. But it is a typical process of executing the photographs in the graphical format. The graph will contain the face features by using the nodes. This will make the detection process so easy. There is a specified network defined as Feed forward neural network. This network will work with the combination of the Gabor wavelet transform to make the analysis success. The face characteristics will consider as the vectors for evaluating the matching technique.