DOC PREVIEW
Memory-based Face Recognition for Visitor Identification

This preview shows page 1-2 out of 7 pages.

Save
View full document
View full document
Premium Document
Do you want full access? Go Premium and unlock all 7 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 7 pages.
Access to all documents
Download any document
Ad free experience
Premium Document
Do you want full access? Go Premium and unlock all 7 pages.
Access to all documents
Download any document
Ad free experience

Unformatted text preview:

Appears in: Proceedings of International Conference on Automatic Face and Gesture Recognition, 2000Memory-based Face Recognition for Visitor IdentificationTerence Sim1Rahul Sukthankar1,2Matthew Mullin2Shumeet Baluja1,21The Robotics Institute2Just ResearchCarnegie Mellon Univ. 4616 Henry StreetPittsburgh, PA 15213 Pittsburgh, PA 15213{tsim,rahuls,mdm,baluja}@justresearch.comAbstractWe show that a simple, memory-based technique forappearance-based face recognition, motivated by the real-world task of visitor identification, can outperform moresophisticated algorithms that use Principal ComponentsAnalysis (PCA) and neural networks. This technique isclosely related to correlation templates; however, we showthat the use of novel similarity measures greatly improvesperformance. We also show that augmenting the memorybase with additional, synthetic face images results in fur-ther improvements in performance. Results of extensiveempirical testing on two standard face recognition datasetsare presented, and direct comparisons with published workshow that our algorithm achieves comparable (or superior)results. Our system is incorporated into an automated vis-itor identification system that has been operating success-fully in an outdoor environment since January 1999.1. IntroductionThe problem of visitor identification consists of the fol-lowing: a security camera monitors the front door of abuilding, acquiring images of people as they enter; an auto-mated system extracts faces from these images and quicklyidentifies them using a database of known individuals. Thesystem must easily adapt as people are added or removedfrom its database, and the system must be able to recognizeindividuals in near-frontal photographs. This paper focuseson the face recognition technology that is required to ad-dress this real-world task.Face recognition has been actively studied [7, 12], par-ticularly over the last few years [9]. The research efforthas focused on the subproblem of frontal face recognition,with limited variance in illumination and facial expression.In this domain, techniques based on Principal ComponentsAnalysis (PCA) [10] popularly termed eigenfaces [26, 16],have demonstrated excellent performance. This paper in-troduces a simple, memory-based algorithm for face recog-nition, termed ARENA, that satisfies the requirements out-lined above and also significantly outperforms PCA-basedmethods on two standard face recognition datasets.2. Image Datasets and PreprocessingOur results use human face images from two standarddatasets: Olivetti-Oracle Research Lab (ORL) [22] andFERET [17, 19]. ORL consists of 400 frontal faces: 10tightly-cropped images of 40 individuals with only minorvariations in pose (±20◦), illumination and facial expres-sion. The faces are consistently positioned in the imageframe, and very little background is visible. FERET con-tains over 1100 faces; however many of them are unsuitablefor our experiments since they are partial or full profiles, orthe individuals were only photographed twice. Therefore,from FERET, we selected the subset of images that satisfiedthe following two constraints: (1) near-frontal poses; (2)images of individuals with more than five such images (ourtests require several images for each person). The resulting275 images consist of 40 individuals, with greater variationin pose and lighting than in the ORL dataset. For instance,many of these images were taken over different days anddisplay significant differences in hairstyles, eyewear, andillumination. Unlike the ORL images, the FERET faces areof non-uniform size and do not always appear in the samelocation of the image. We use the FERET images as pro-vided to explore the potential limitations of our template-based face recognition technique. Figure 1 shows two im-ages for each of two individuals from the two datasets.Figure 1. Top row: Two sample images each, of two subjects from ORL (left), and FERET (right).Note the difference in facial orientation, expression and accessories between the two images of thesame individual. FERET images tend to exhibit greater variation in appearance (including hairstyle andclothing). Bottom row: the corresponding ARENA reduced-resolution images (16 × 16 pixels).3. The ARENA Face Recognition AlgorithmARENA is a memory-based [1] algorithm that employsreduced-resolution images (16 × 16) and the L∗0similaritymeasure (described below). The reduced-resolution imagesare created by simply averaging over non-overlapping rect-angular regions in the image. The distance from the queryimage to each of the stored images in the database is com-puted, and the label of the best match is returned.3.1. L∗psimilarity measuresOur results show that the obvious choice for ARENA’ssimilarity measure, the Euclidian distance, performs poorly.In this section we present alternatives. The Lpnorm is de-fined as: Lp(~a) ≡ (P|ai|p)1p. Thus, the Euclidian distanceis simply: L2(~x−~y). Note that since we are not interested inthe actual distances, but only in the ordering, we can equiv-alently employ the similarity measure L∗p(~a) ≡ (P|ai|p).Robust statistics literature shows that L∗2, despite its con-venient analytic properties, overly penalizes outliers [11].For this reason, the L∗1similarity measure is often used innoisy environments. For ARENA, we have explored severalL∗psimilarity measures (see Figure 2). The L∗0similaritymeasure is defined as L∗0(~a) ≡ limp→0+L∗p(~a). Intuitively,L∗0(~x − ~y) counts the number of components in ~x and ~ythat differ in value. Our experiments indicate that the bestperformance on this task is achieved with p ≤ 1.In our application, each reduced-resolution image is con-verted into a vector, ~x, where each pixel in the image is rep-resented as a component of the vector. In practice, sinceindividual pixel intensities are noisy, we relax the definitionof L∗0to be:L∗0(~x − ~y) ≡X|xi−yi|>δ1where δ is a threshold, such that pixels whose intensitiesdiffer by less than δ are considered equivalent.4. Principal Components Analysis (PCA)The most widely used baseline for face recognition,eigenfaces [26, 16] employs Principal Components Anal-ysis (PCA), which is based on the discrete Karhunen-Lo`eve(K-L), or Hotelling Transform [10], is the optimal linearmethod for reducing redundancy, in the least mean squaredreconstruction error sense. Points in Rdare projected intoRm, (where m ≤ d, and typically m  d). PCA has be-come popular for


Memory-based Face Recognition for Visitor Identification

Download Memory-based Face Recognition for Visitor Identification
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...
Login

Join to view Memory-based Face Recognition for Visitor Identification and access 3M+ class-specific study document.

or
We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view Memory-based Face Recognition for Visitor Identification 2 2 and access 3M+ class-specific study document.

or

By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?