IMPLEMENTATION METHOD VIOLA JONES FOR DETECTION MANY FACES Liza Angriani 1,Abd. Rahman Dayat 2, and Syahril Amin 3 Abstract In this study will be explained about how the Viola Jones, and apply it in a lot of face detection system in real time by utilizing OpenCV libraries. Once the system is completed, will discuss how the detection system ranging from image acquisition, image processing, pattern recognition, and image analysis. Then testing on the character's face can be detected. Training data used in this study consists of two (2) parts, ie positive samples and negative samples. Positive sample is a sample image of human, while a negative image of trees and animals. The data used in this research is an image taken at random from the webcam and internet as much as maximum human image and varian distance. The image is used most small-sized dimensions are 640x480 pixels and most greatest dimension is 1600x1200 pixels. Training data is performed using the OpenCV library. Here's a face detection process scheme using Viola-Jones method Management of Umel Mandiri Jayapura, Papua, Indonesia (e-mail: ril.nocturno@gmail.com). Keywords image recognition, face detection, viola-jones. I. INTRODUCTION I mage processing is a system where the process is done with the input of the original image and also in the form of images. One area that uses image processing which is currently developed a lot of people are biometric applications, ie the field that studies how to be able to identify the unique characteristics that exist in the brain of the human body can determine whether the object seen by the human eye or it is not, face or hands and so on. One of the fields in computer vision that many do is research on face recognition which is the domain of computer vision applications. 1 Commercial application of this has been implemented, but basically this is not a perfect technology research needs to be developed to achieve the desired result. One thing that can be added or used in the development of facial recognition technology is to increase the level of speed and accuracy in detecting faces. Many of the detection system using the Viola Jones [1] as an object detection method. The Viola Jones Methods is known to have the speed and accuracy is quite high because it combines several concepts (Haar features, Integral image, AdaBoost, Cascade Classifier) became a major method for detecting objects. Several studies using the method Viola Liza Angriani 1 is with the Academy of Computer and Information Management of Umel Mandiri Jayapura, Papua, Indonesia (corresponding author s phone: +6281320413078; e-mail: liza.angriani2@gmail.com). Abd. Rahman Dayat 2, was with the Academy of Computer and Information Management of Umel Mandiri Jayapura, Papua, Indonesia (e-mail: aldurra.afgan@live.com). Syahril Amin 3, was with the Academy of Computer and Information 1
And Jones among others distance between the finger to the camera is within 10-15 cm, the position of the finger is detected at the position of 0 and 45 in the study concluded Viola Jones method is able to detect in real time and has the accuracy of the high [2]. Game Tic Tac Toe with Finger Gestures which concluded that the more light around the room, resulting in an excellent finger detection [3]. In addition, this method is used also as a method for tracking an object system [4]. II. IMAGE PROCESSING Image processing is an important part of the underlying variety of real applications, such as pattern recognition, remote sensing by satellite and machine vision [5]. Image processing is a method or technique that can be used to process the image with the image or manipulated into the desired image data to obtain specific information [6]. Digital image obtained from the digitized analog image. Image digitization involves two processes, namely sampling and quantization. Sampling shows the number of pixels / block to define an image. Quantization shows the number of degrees of value at each pixel (indicates the number of bits in a digital image, black / white with 2 bits, with 8-bit grayscale, with 24- bit true color). Image processing is done to improve the quality of the image to make it easier to interpret by human / computer. Input and output are also the image of the image, but with a better quality than the input image. Image processing operations associated in face detection: grayscaling, neighborhood operations, thresholding, histogram equalization, resizing III. VIOLA-JONES METHODS Viola- Jones method is divided into four main components: Haar Like Feature, Integral Image, Adaptive Boosting and Cascade of Classifier. Haar Feature Haar feature is a feature that is used to detect an object. Features Haar feature based on Haar wavelet [1]. Haar wavelet is a single square wave having a high interval and a lower interval. Furthermore, combinations of box used for the detection of visual objects better. Each Haar-like feature consists of a combined box - black and white box as follows: Edge features (a) (b) (c) (d) 2
Lines Features 1 2 3 1 3 6 4 5 6 5 12 21 (a) (b) (c) (d) (e) (f) (g) Center Surround (h) 7 8 9 12 27 45 (a) Orginal Image (b) Integral Image Fig. 3 Example of calculation of the integral image (a) (b) Fig. 1 Various Kinds of Variations Featured on Haar [7] In the Viola-Jones method, Haar feature known as Haar- Like feature. Haar-Like feature will process the image in boxes containing several pixels of an image. Determination of Haar feature is determined by subtracting the total pixels in the dark areas of the total pixels in the bright areas. Can be written with the formula: To obtain a faster computation can use the integral image. A. Integral Images Integral images is a method for calculating the box Haar-like features. This method is used to calculate quickly without calculating the overall value of the existing pixels in the feature. Integral images at location x, y consists of the value of the number of pixels above and left of location x, y. Fig. 2 Integral Images According to Viola and Jones Integral images at the location x, y, containing the number of the top to the left of x, y calculation using the following formula: To calculate the integral image by ignoring the original image, using a pair the following formula: Where : i(x,y) ii(x,y) s(s,y) s(x,-1) = 0 ii(-1,y) = 0 P(x,y) = Integral Image is original image : Orginal Image : integral image : The cumulative number of rows Fig. 4 The integral image calculation Can be seen in the picture above is based on the calculation of the number of pixels in a square D, can be calculated with reference point array. The value of the integral image at location 1 is the sum of the rectangle A. Value at location 2 is A + B, the value at location 3 is A + C, and at the location 4 is A + B + C + D. summation role in D could be count like 4 + 1- (2 + 3). Iterative search process is done from the top is now on the image to the bottom right corner of the image. B. Adaptive Boost (AdaBoost) Most of boosting algorithms follow a design that adds a weak learner into a strong learner. Weak learner is a classifier that has little correlation with the actual classification. While strong learner is a classifier which has a strong correlation with the actual classification [8]. At each iteration, a weak classfier learn from one exercise, then the weak learner will be added to the strong learner. Adaboost is an abbreviation of Adaptive Boosting formulated by Yoav Freund and Robert Schapire in 1995. The AdaBoost algorithm is used to perform election-features in large quantities by simply selecting certain features. AdaBoost algorithm is an algorithm that is able to adapt with less weak classifier. According to Viola and Jones [4] one practical method to solve it is to limit the weak learner to the classification of functions, each of which relies on a single feature. For each feature determining thresholds weak learner optimal classification function. C. Casade Classifier A C B 1 2 D 3 4 Casscade Classifier combination is the final step in the Viola- Jones method. By combining the classification in a multi-storey structure, the speed of the process can be increased is by focusing on the areas in the image that is likely to be. This is done to determine where the location of the object you are looking at an image. 3
All sub -windows 1 F Fig. 5 Implementation of Face Detection with two faces T T T 2 3 F F Reject Sub window Futher Processing Fig. 4 Flow cascade classifier [1] At the first level of classification, each sub-image will be classified using the feature. The results of this first classification is T (True) for images that meet certain Haar features and F (False) if not. This classification will be leaving roughly 50% sub-image to be classified in the second stage. The results of the second classification is T (True) for images that meet the integral image and F (False) if not. Along with increasing levels of classification, it would require more specific terms that are used feature becomes more. Sub-image who escapes classification number would be reduced up to a total of about 2% [4]. The results of the latter classification is T (True) for images that meet the AdaBoost and F (False) if not.. IV. CONSTRUCT AND IMPLEMENTATION The samples used were taken from the images on the internet, which consists of the positive samples and negative samples. Positive samples is that the pattern of the human face, while the negative samples is any image that is not a human face pattern. Each sample using 50 different images with low resolution of 640x480 pixels up to 1600x1200 pixels. After the system design has been completed, the system is ready for the construction and implementation. First of all the software design is constructed under desktop platform using OpenCV. The system environment itself we used Microsoft Visual Studio 2012. The construction step consists of graphical user interface (GUI). Fig. 6 Implementation of Face Detection with four faces Experiments have been performed to test the proposed system and measure the accuracy of the system. The system is designed in the Microsoft Visual Studio 2012, ADT 21.1, Library OpenCV to face recognition. The image to the color system of the input to the image with a size of 640 x 480. Test images taken under different distance conditions with 10 peoples in real time. The test results are shown in table I. Indicated that the accuracy for the character segmentation is 99,6%, 98%, 97,7% and 97,6% is the percentage of unit recognition accuracy. Overall system performance can be defined as the product level of accuracy of all units (Detection face recognition). TABLE I TABLE RESULT No. Distance of FR System Percentage of Accuracy (%) 1. 1 meter 99,6% 2 2 meters 98% 3. 3 meters 97,7% 4. 4 meters 97,6% V. CONCLUSION This paper has been discussed on design and implementation of the introduction of Many Face Recognition (MFR) using Viola-Jones Methods with Haar Featured for the detection face pattern on desktop platform. In conclusion, we have proposed the design of the MFR can be implemented on desktop platform in real time. The system is designed for face and identification system in real time have been tested. Finally, the results show that the proposed system to reach 99.6% accuracy detection in within close and 97,5% for few far distance. In addition, the proposed system can be can be developed for face recognition for tracking people in a public space in the future. 4
REFERENCES [1] Viola, P., Jones, M. J.,(2001) Rapid Object Detection Using A Boosted Cascade of Simple Features, IEEE Conference on Computer Vision and Pattern Recognition, Jauai, Hawaii,. [2] Chandra, Devy. Prajnagaja, Nagarjuna. Nugroho, Lintang Agung, (2011), Studi Pendeteksian Wajah dengan Metode Viola Jones, BINUS. [3] Nugraha, Raditya, (2011), Game Tic Tac Toe dengan Gerakan Jari Menggunakan Metode Viola And Jones, Politeknik Elektronika Negeri Surabaya, ITS. [4] Arihutomo, Mukhlas, (2010), Rancang Bangun Sistem Penjejakan Objek Menggunakan Metode Viola Jones Untuk Aplikasi EyeBot, ITS. [5] Abduk Kadir, Adhi Susanto, 2013. Teori dan Aplikasi Pengolahan Citra. Penerbit ANDI, Yogyakarta. [6] Murni, Aniati, (1992). Pengantar Pengolahan Citra, PT Elex Media Kompuindo, Jakarta [7] Rainer Lienhart and Jochen Maydt, (2002). An Extended Set of Haar-like Features for Rapid Object Detection, Intel Labs, Intel Corporation. [8] Andoko, (2007). Perancangan Pro gram Simulasi Deteksi Wajah Dengan Support Vector M achines-viola Jones. Liza Angriani is a Lecturer AMIK Umel Mandiri Jayapura. currently continuing graduate study at Postgraduate STMIK Handayani Makassar. Abd. Rahman Dayat is a Lecturer AMIK Umel Mandiri Jayapura. Currently continuing graduate study at Postgraduate STMIK Handayani Makassar. Syahril Amin is a Lecturer AMIK Umel Mandiri Jayapura. Currently continuing graduate study at Postgraduate STMIK Handayani Makassar. 5