114IEEE TRANSACTIONS ON NEURAL NETWORKS,VOL.8,NO.1,JANUARY1997 Face Recognition/Detection by ProbabilisticDecision-Based Neural NetworkShang-Hung Lin,Sun-Yuan Kung,Fellow,IEEE,and Long-Ji LinAbstract—This paper proposes a face recognition system based on probabilistic decision-based neural networks(PDBNN).With technological advance on microelectronic and vision system,high performance automatic techniques on biometric recognition are now becoming economically feasible.Among all the biomet-ric identification methods,face recognition has attracted much attention in recent years because it has potential to be most non-intrusive and user-friendly.The PDBNN face recognition system consists of three modules:First,a face detectorfinds the location of a human face in an image.Then an eye localizer determines the positions of both eyes in order to generate meaningful feature vectors.The facial region proposed contains eyebrows,eyes, and nose,but excluding mouth.(Eye-glasses will be allowed.) Lastly,the third module is a face recognizer.The PDBNN can be effectively applied to all the three modules.It adopts a hier-archical network structures with nonlinear basis functions and a competitive credit-assignment scheme.The paper demonstrates a successful application of PDBNN to face recognition applications on two public(FERET and ORL)and one in-house(SCR) databases.Regarding the performance,experimental results on three different databases such as recognition accuracies as well as false rejection and false acceptance rates are elaborated in Section IV-D and V.As to the processing speed,the whole recog-nition process(including PDBNN processing for eye localization, feature extraction,and classification)consumes approximately one second on Sparc10,without using hardware accelerator or co-processor.Index Terms—Decision-based neural network(DBNN),prob-abilistic DBNN,face detection,eye localization,virtual pattern generation,positive/negative training sets,hierarchical fusion, face recognition system.I.I NTRODUCTIONW ITH its emerging applications to secure access con-trol,financial transactions,etc.,biometric recognition systems(e.g.,face,palm,finger print)have recently taken on a new importance.With technological advance on mi-croelectronic and vision system,high performance automatic techniques on biometric recognition are now becoming eco-nomically feasible[3].Among all the biometric identification methods,face recognition has attracted much attention in recent years because it has potential to be most nonintrusive and user-friendly.In this paper we propose an integrated face recognition system for security/surveillance purposes.This system involves three major tasks:1)human face detection from still images and video sequences;2)eye localization;Manuscript received February15,1996;revised June19,1996.S.-H.Lin and S.-Y.Kung are with the Department of Electrical Engineering, Princeton University,Princeton,NJ08540USA.L.-J.Lin is with Siemens SCR Inc.,Princeton,NJ08540USA. Publisher Item Identifier S1045-9227(97)00233-6.and3)face recognition/rejection.Several different techniques have been proposed over the last20years to tackle the above three problems.A brief review of those research efforts is as follows.The key issue and difficulty in face detection is to account for a wide range of variations in facial images.There exist several approaches for dealing with such variations,includ-ing:1)spatial image invariants;2)correlation template(s); and3)view-based eigen-spaces,etc.Some schemes exploit some common and unique spatial image relationships existing among all face patterns,called image invariants.One such image-invariant is the local ordinal structure of brightness distribution between different parts of a human face[4].The schemes check these invariants for positive occurrences at all image locations.Yet the correlation template approach computes a similarity measurement between afixed target pattern and candidate image locations.If the correlation ex-ceeds a certain threshold,then a face is confirmed,i.e., detected.Some face detection approaches use a bank of correlation templates to detect major facial subfeatures in the image[5],[6].However,the set of allowable facial patterns is probably too large to be adequately represented byfixed template or templates.A closely related approach is that by means of view-based eigenspaces[7].The approach assumes that the set of all possible face patterns occupies a small and parameterized subspace,derived from the original high-dimensional input image space.Typically,the approach approximates the subspace of face patterns using data clusters and their principal components from one or more example sets of face images.An image pattern is classified as“a face”if its distance to the clusters is below a certain threshold, according to an appropriate distance metric.An advantage of eigenspace approaches is that it is possible to deal with occluded situations.Several image reconstruction and pose estimation examples based on eigenface method are included in[7].However,there is no experimental result showing the detection rate of occluded faces in the paper.Also, this approach has so far only demonstrated to be working in uncluttered background.Another eigenspace method is reported in[8].this method estimates the likelihood function of the face class from the principal face subspace and then decide whether the input image pattern is a face or not based on its likelihood function value.Again,[8]only demonstrates its performance on the database containing faces in uncluttered background(the ARPA FERET database).This method is similar to our approach,whose objective is also to estimate1045–9227/97$10.00©1997IEEEAuthorized licensed use limited to: University of Science and Technology of China. Downloaded on October 16, 2009 at 12:35 from IEEE Xplore. Restrictions apply.
本文来源：https://www.2haoxitong.net/k/doc/09192704cc175527072208e2.html
