Documents

Face and Eye Detection by CNN Algs

Categories
Published
of 15
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Description
Face description
Transcript
  Journal of VLSI Signal Processing 23, 497–511 (1999)c  1999 Kluwer Academic Publishers. Manufactured in The Netherlands. Face and Eye Detection by CNN Algorithms DAVID BALYA ∗ AND TAM´AS ROSKA  Analogical and Neural Computing Laboratory, Computer and Automation Research Institute, L´ agym´ anyosi u. 11, Budapest, H-1111, Hungary Abstract.  A novel approach to critical parts of face detection problems is given, based on analogic cellular neuralnetwork (CNN) algorithms. The proposed CNN algorithms find and help to normalize human faces effectivelywhile their time requirement is a fraction of the previously used methods. The algorithm starts with the detectionof heads on color pictures using deviations in color and structure of the human face and that of the background. Bynormalizing the distance and position of the reference points, all faces should be transformed into the same sizeand position. For normalization, eyes serve as points of reference. Other CNN algorithm finds the eyes on anygrayscale image by searching characteristic features of the eyes and eye sockets. Tests made on a standard databaseshow that the algorithm works very fast and it is reliable. 1. Introduction This paper deals with a novel approach to the facedetection problem, based on cellular neural networks(CNN) processing. It describes a completely newmethod for the localization and normalization of faces,which is a critical step of this complex task but hardlyever discussed in the literature. The proposed CNNalgorithm is capable of finding and helping normalizehuman faces effectively while its time requirement is afraction of the previously used methods (see Table 1).Recognition of faces is a remarkable example forthe ability of humans to perform complex visual tasks.Even animals can solve this task. Pigeons find theirmates easily though they look very similar to us. Ourrecognitionabilitydevelopsoverseveralyearsofchild-hood. This is important in several aspects of our sociallife such as estimating the facial-expression of peoplewe interact with. It has probably played an importantrole in the course of evolution.Face recognition, a problem that has been consid-ered a challenge since the very first days of computervision, has experienced a revival recently, after nearlytwenty years of silence. Finding proper solution to theproblem was one of the goals of the fifth generation ar-tificialintelligenceproject. Severalmethodshavebeen ∗ D. Balya is also with the Technical University of Budapest. proposed but none of them are considered to be perfectin all aspects such as accuracy, speed etc.Inourapproachthecomplextaskisdividedintosub-tasks. The literature deals mainly with the representa-tion of faces i.e. the phase of the problem that comesafter finding and normalizing the face qualitatively aswell as geometrically.The first step, however simple it may seem, con-stitutes in reality a difficult and sophisticated pre-processing sub-task. Most known systems spend morethan half of the total computation time on this step,which is due to the quality of the whole process de-pending on the success of the detection. The presentpaper utilizes the potential of a new tool, the CellularNeural Network (CNN) computer, in which an arrayof programmable analog processing cells perform par-allel computation on corresponding pixel values of agiven image [1–5]. Templates control functioning of a CNN-UM (Cellular Neural Network and UniversalMachine) and the programming of a CNN is done byorganizing the series of templates [6].The human face as input information could be usedby various systems [7]. Biometrics identification sys-tems present an obvious example because human facesare straightforward and comfortable objects for identi-fication. Security systems, e.g. ATMs, could use facesas keys which can not be lost and are very difficultto forge. In working with image databases, such as  498  Balya and Roska Table 1 . Comparing some eye detection algorithms.Tool Accuracy Speed RemarksPDBNN [8] 96.4% 500 ms SUN Sparc20 It needs many images for trainingDT, features [23] 96%  < 1 sec Part of a face recognition systemMIP [24] 95% 91 sec SGI INDY Find noses tooTemplates [25] 94% 50 ms TMS320C40 Poorer on static imagesNeural Nets [26] 93% 120 sec SUN Sparc20 Face detectionCNN 98% 3 ms CNN chip See Section 1 and 6 television and video, searching for given visual objectsis an inherent task. The actual identification task oftenrequires the witness to scrutinize thousands of imagesincluding line drawings or sketches. Obviously, this istoomuchforasuccessfulsearch. However, theimagescan effectively be used for the control of alertness.The algorithm described in the paper starts withthe detection of the head on a colored picture usingdeviations in color and structure of the human faceand that of the background. Reference points areneeded for normalization of the face. The eyes werefound the most reliable for this purpose. By normal-izing the distance and position of reference points, allfaces can be transformed into the same dimension andposition.The proposed CNN algorithm finds the eyes on anygrayscale image by searching for some characteristicfeatures of the eyes and eye sockets. Test made on astandarddatabase(15randomlyselectedfacesfromtheFERET face database and 15 other images) shows thatthe algorithm works very fast (see Section 6) and thatit is reliable. Its error (non-detection) is a few percent(in one case the algorithm did not find the eyes, thoughin several cases it found more objects than the twoeyes).In this paper, we introduce the following topics: inSection two we discuss the face recognition steps. InSection three the head-detection problem is describedand a CNN algorithm is proposed as solution. Thefourth section talks about face normalization. Thefifth section is devoted to the eye localization. Sectionsix describes the details of the developed eye-finderCNN algorithm, while conclusion is drawn in Sectionseven. 2. Face Recognition Overview Face recognition is described as follows identify theperson(s) in a picture that is given on arbitrary media.Pictures can come from any scanned in photograph ordigitalgraphicsoncomputeraswellasashotofavideofilm. Face recognition is a complex task, which can bedivided into several steps (Fig. 1).In order to recognize objects, one must successfullylocatetheobjectsandextractusefulfeaturesfromthem.For face recognition, the first step is to detect the pres-ence of human head. This detection must not takea long time; it should use only one or two proper-ties which can be checked fast. In video processingit means real-time because this is only the first filteringoftheframes. Themethodshouldhavelowfalserejec-tionrate(FRR).Wegiveasimplecolorbaseddetectionalgorithm in section three.Head localization is the next step in the procedure.Forlocalizingthehumanheadonecouldusemorecom-plicated features. This is a critical step in automatedface recognition. CNN algorithms (see Section 6) canaid the complex feature extraction.Further processing begins with a complex task,termed as capturing the face. The head position in-formation is known from the previous step but the faceshould be normalized. The head is transformed in or-der to create canonical images, in which the face isscaled and rotated to a standard size. The process upuntiltothisstagerequiresabouthalfofthecomputationtime of the whole face recognition [8]. It needs spaceand processing power in a traditional way, therefore if this normalization can be speeded up, the time of thewhole procedure will be reduced proportionally. TheCNN is an appropriate tool for accomplishing the task (see Section 4).It should be determined what features are the bestto represent the object. Two criteria are used to se-lect a good feature vector: it must convey enoughinformation for distinguishing the object from oth-ers, and it must be as invariant as possible. Com-puter vision research groups have developed manydifferent methods for face representation. Severalmethods of different complexity are described in theliterature ranging from PC computation power to  Face and Eye Detection by CNN Algorithms 499 Figure 1 . Stages of face recognition. Face recognition as a complex activity can be divided into several steps from detection of presence todatabase matching. The literature deals mainly with the representation and identification of faces.  500  Balya and Roska supercomputing and from traditional methods [9] tosoft computing paradigm [10] and statistics [11].The feature vector can then be used to recognizea face from a given database. The database containsfeature vectors of the people stored in it and/or theirstandard picture(s). The process identifies the personwhose feature vector and that of the agent face has aminimum distance [12] if the difference is lower than agiven threshold. The goal is to minimum false accep-tance (FAR)  and   false rejection (FRR). 3. Head Detection Theheaddetectionproblemcanbedescribedasfindingthe location and size of every human face within theimage. The head can be detected if its representativeregion is observed [13]. It is obvious that the wholefrontal facial area contains information for recogni-tion. The job of deciding which part of the object isthe representative region is done before the detectionand recognition processes begin. The developed CNNalgorithmisbasedonnaturalhumancolorinformation.The complexity of head localization srcinates fromthe nature of humans and their environment. ã  Thesizeofthefaceinthescenecanbediverse,there-fore we need size invariant procedure to mask theheads. ã  The background clutter causes difficulty especiallyif the head of the person is relatively small in theimage. The texture of the background and clothesmay lead to false localization as well. ã  Diverseheadposes, inclinationandrotationproducevery different visual objects (e.g. profile versusfrontal views). This problem can be turned to ourbenefit if the image is from a live video. Then wecan test liveliness with different face poses so theface recognition system will be safer. Intruders arenot able to use static pictures for identifying them-selves. 3.1. Human Head Features The human head has some basic features. Many fea-tures such as bilateral symmetry, elliptical outline, notuniform field (as some background) are characteristicto humans.First, we need to localize the person in the picture.In this task we can use color information. Human skinis never blue or green therefore we may assume thatorangeisinthecenterandthatthereisabroadspectrumaroundthiscenter,whichcanaccommodateyellowandwhite as well as light brown. Color recognition is afast method for detecting the presence of human head[14, 15]. 3.2. A Simple Color Head LocalizationCNN Algorithm Detection of some head features can be done simplyusing CNN algorithms and it also solves most of theproblems described previously. The CNN can performsize invariant algorithms, therefore it can extract headsof very different sizes. The background can usually beremoved with skin color information. Clothes rarelyhave the same color as the face color.At first, we apply color filter which involves some threshold   template and logic instructions [16]. Thenthe gradient field is computed and the smallest objectsare erased using the  small killer   template, because thehuman skin is non-uniform. After running of the al-gorithm, we get a mask of the color picture which in-dicates the possible head fields (Fig. 2). All the tem-plates used in this task can be found in the CSL (CNNtemplate library) [17]. 4. Face Normalization The possible head positions are known from the headdetection step. The face capturing stage normalizesthose faces (Fig. 3). Producing faces of standard sizeanddirectionisanimportantandtime-consumingstep.We were able to construct a CNN algorithm which sig-nificantly helps to solve this task.Reference points should be found in the face to nor-malize it. One of the most fundamental references iseye position, because eyes are significantly differentfrom other parts of the face and every human has gottwo of them. Once the eye positions are determined,all other important facial features, such as positions of nose and mouth, can easily be determined. Translationdependency can be eliminated by setting the srcin of co-ordinates to a point that can be detected with suf-ficient accuracy in each image (e.g. eyes). We devel-oped a CNN algorithm for localizing the eye areas (seeSection 6).Additional features should be extracted from the de-tected head. Detection of the nose line and that of 
Search
Similar documents
View more...
Tags
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks