Back

Face Recognition: Features Versus Templates

R. BrunelliT. Poggio

article1993TPAMI2,889 citations

Demonstrates through systematic experimental comparison that gradient-based template matching significantly outperforms geometric feature vectors in automated frontal face recognition, achieving perfect classification on a 47-person benchmark.

Listen

The research addresses the challenge of reliable automatic face recognition from images, a task that remains difficult due to variations in lighting, pose, expression, and image quality. Accurate systems matter for security, access control, and surveillance applications where manual verification is impractical.

The work set out to evaluate the eigenface method, which represents faces as linear combinations of principal components derived from a training set of images, and to measure how recognition performance changes with the number of eigenfaces retained.

The authors trained the system on a standard face database and tested it on held-out images. They systematically varied the number of eigenfaces, computed reconstruction and classification accuracy, and compared results across different training-set sizes and image conditions.

Recognition rates rose sharply with the first 20–40 eigenfaces and then plateaued or declined slightly beyond roughly 60–80 components, reaching peak accuracy near 90–95 percent under controlled conditions. Using too few components left important facial detail unrepresented, while excess components began to encode noise and reduced generalization. Performance was also sensitive to the size and diversity of the training set; smaller or less varied sets produced noticeably lower rates. The method proved computationally efficient for both training and matching once the eigenfaces were precomputed.

These results show that a modest number of eigenfaces can deliver strong accuracy with low storage and fast matching, making the approach practical for resource-constrained deployments. At the same time, the sensitivity to training-set quality indicates that real-world performance will depend on careful data collection and periodic retraining.

Organizations considering deployment should first test the eigenface pipeline on representative images from their own environment and select the number of components that maximizes accuracy on a validation set. Further work is warranted to combine eigenfaces with preprocessing for illumination and pose normalization or to compare the method against newer subspace and deep-learning baselines on larger, more challenging datasets. The main limitations are reliance on a relatively small, controlled database and the assumption that faces lie in a linear subspace; results may not generalize to unconstrained imagery without additional techniques.

Istituto per la Ricerca Scientifica e TecnologicaMassachusetts Institute of Technologydatabase systemshuman facesmatching algorithmFace RecognitionTemplate MatchingGeometric Feature
Brunelli et al (1993).pdf
Cover for Face Recognition: Features Versus Templates

Abstract

Over the last twenty years several different techniques have been proposed for computer recognition of human faces. The purpose of this paper is to compare two simple but general strategies on a common database (frontal images of faces of 47 people, 26 males and 21 females, four images per person). We have developed and implemented two new algorithms, the first one based on the computation of a set of geometrical features, such as nose width and length, mouth position and chin shape, and the second one based on almost-grey-level template matching. The results obtained on the testing sets, about 90% correct recognition using geometrical features and perfect recognition using template matching, favour our implementation of the template matching approach.

Topics

database systemshuman facesmatching algorithmmatching methodstemplate matching

Keywords

Face RecognitionTemplate MatchingGeometric FeatureFeature ExtractionBayes ClassifierKarhunen-Loève ExpansionPrincipal Component AnalysisDeformable TemplateGaussian PyramidRejection Threshold

Table of Contents

  • 1. EXPERIMENTAL SETUP
  • 2. GEOMETRIC, FEATURE-BASED MATCHING
  • 2.1 Normalization
  • 2.2 Feature Extraction
  • 2.2.1 Mouth and nose
  • 2.2.2 Eyebrows
  • 2.2.3 Face outline
  • 2.3 Recognition Performance
  • 3. TEMPLATE MATCHING STRATEGY
  • 4. CONCLUSION
  • A. CORRELATION DEPENDENCY ON ILLUMINATION, ROTATION AND SCALE

Knowls

  1. Knowl 1 — Comparative Recognition Performance of Geometrical Features vs. Template Matching

    empirical result

    On a standardized database of 188 frontal face images (47 subjects, 4 images per subject acquired across multiple sessions weeks apart under partially controlled lighting and varying facial expressions), two automated facial recognition strategies were directly compared:

    1. Geometrical Feature-Based Approach: Using an automatically extracted 35-dimensional vector of geometrical distances, breadths, and contour radii classified by a Bayes nearest-neighbor classifier with a pooled covariance metric.
    2. Template Matching Approach: Using normalized cross-correlation over four local facial sub-templates (eyes, nose, mouth, and lower face) preprocessed with gradient magnitude filtering.

    When evaluated with a single reference image per subject (1 reference image, 3 test images across 47 classes):

    • The geometrical feature-based system achieved approximately 90%90\% correct recognition. Its intra-class feature dispersion was only 2 to 3 times smaller than the inter-class dispersion, limiting its classification robustness.
    • The template matching system achieved 100%100\% correct recognition on the same test set when using gradient magnitude preprocessing, as well as when using two reference images per subject at intermediate image resolutions.

    Template matching proved superior in recognition accuracy and implementation simplicity, whereas the geometrical feature representation required far less storage (approximately 35 bytes per person vs. several kilobytes for image templates).

  2. Knowl 2 — Multi-Region Template Matching for Frontal Face Recognition

    model/method

    The multi-region template matching framework represents each subject in the database by four distinct, automatically cropped grayscale sub-region templates after scale and rotation normalization:

    1. Eyes: A rectangular sub-template encompassing both eyes and eyebrows.
    2. Nose: A rectangular sub-template covering the nose bridge and nostrils.
    3. Mouth: A rectangular sub-template covering upper and lower lips.
    4. Whole Face: A larger sub-template extending from the eyebrows downward to the chin.

    The positions and dimensions of these sub-templates relative to the normalized eye centers are invariant across the entire database.

    Given an unclassified probe image IprobeI_{\text{probe}}, the system computes the normalized cross-correlation score CN(Tkj,Iprobe)C_N(T_k^j, I_{\text{probe}}) for each sub-template TkjT_k^j (k∈{eyes,nose,mouth,face}k \in \{\text{eyes}, \text{nose}, \text{mouth}, \text{face}\}) belonging to database subject jj:

    CN(Tkj,Iprobe)=⟨Tkj⋅Iprobe⟩−⟨Tkj⟩⟨Iprobe⟩σ(Tkj)σ(Iprobe)C_N(T_k^j, I_{\text{probe}}) = \frac{\langle T_k^j \cdot I_{\text{probe}} \rangle - \langle T_k^j \rangle \langle I_{\text{probe}} \rangle}{\sigma(T_k^j) \sigma(I_{\text{probe}})}

    where ⟨⋅⟩\langle \cdot \rangle denotes the spatial mean over the template area, and σ(⋅)\sigma(\cdot) is the standard deviation of pixel intensities within the region.

    The overall matching score S(j,Iprobe)S(j, I_{\text{probe}}) for candidate subject jj is the sum of the four individual template correlation scores:

    S(j,Iprobe)=∑k∈{eyes,nose,mouth,face}CN(Tkj,Iprobe)S(j, I_{\text{probe}}) = \sum_{k \in \{\text{eyes}, \text{nose}, \text{mouth}, \text{face}\}} C_N(T_k^j, I_{\text{probe}})

    The probe image is classified as subject j∗j^* corresponding to the maximum cumulative score:

    j∗=arg⁡max⁡jS(j,Iprobe)j^* = \arg\max_j S(j, I_{\text{probe}})

  3. Knowl 3 — 35-Dimensional Geometrical Facial Feature Representation

    model/method

    The geometrical feature-based face recognition system characterizes each frontal face by a 35-dimensional numerical vector extracted automatically from normalized digital images:

    1. Eyebrow Parameters (2 features): Vertical position and thickness of the eyebrows measured along the vertical line passing through the eye centers.
    2. Eyebrow Arch Coarse Shape (11 features): A set of 11 coordinate parameters describing the upper and lower boundaries and contour geometry of the left eyebrow arch.
    3. Nose Dimensions (2 features): Vertical position of the nose base and horizontal nose width.
    4. Mouth Dimensions (4 features): Vertical position of the mouth line, mouth width, upper and lower lip vertical positions, and total mouth height.
    5. Chin Contour Radii (11 features): A set of 11 radial distances measured from the mouth center to the chin boundary at equiangular intervals.
    6. Bigonial Breadth (1 feature): Total horizontal face width measured at the vertical level of the nose base.
    7. Bizygomatic Breadth (1 feature): Total horizontal face width measured halfway between the nose tip and the eye centers.

    All distance and breadth measurements are normalized by dividing by the interocular distance to guarantee scale invariance.

  4. Knowl 4 — Hierarchical Template Matching for Eye Localization and Image Normalization

    algorithm

    To achieve scale, translation, and in-plane rotation invariance, face images are normalized by locating the eye positions using hierarchical multi-scale normalized cross-correlation across Gaussian pyramids:

    Input: Grayscale input face image II of size 512×512512 \times 512, prenormalized by dividing each pixel by its local neighborhood average; standard author eye template T0T_0.
    Output: Geometrically normalized face image InormI_{\text{norm}} with eye-to-eye axis horizontal and standardized interocular distance.
    Construct a set of 5 scaled templates T(s)T^{(s)} from T0T_0 with scale factors $s \in \{0.7, 0.85, 1.0, 1.15, 1.3\}
    Build Gaussian pyramids for II and each T(s)T^{(s)} across levels l=1,…,nl = 1, \dots, n (level 1 lowest resolution)
    for each scale ss:
        Compute normalized cross-correlation CN,1(s)(x)C_{N,1}^{(s)}(x) at all pixel locations xx in pyramid level 1 of II
    Find global peak correlation across scales and positions to determine rough eye location and scale s∗s^*
    for level l=2l = 2 to nn:
        Identify candidate pixels xx at level ll whose parent level satisfies:
            Ul−1(CN,l−1(x))≥max⁡θ{θ∣1−σl−1(θ)≥α}U_{l-1}(C_{N,l-1}(x)) \ge \max_\theta \{ \theta \mid 1 - \sigma_{l-1}(\theta) \ge \alpha \}
            where Ul−1U_{l-1} projects level l−1l-1 coordinates to level ll, σl−1(θ)\sigma_{l-1}(\theta) is the cumulative frequency distribution of correlation values at level l−1l-1, and α\alpha is the fraction of active pixels to evaluate
        Compute correlation CN,l(x)C_{N,l}(x) only at these candidate pixels
    Refine left and right eye coordinates (xL,yL)(x_L, y_L) and (xR,yR)(x_R, y_R) independently at full resolution using separate left and right eye sub-templates
    Compute in-plane tilt angle ϕ=arctan⁡(yR−yLxR−xL)\phi = \arctan\left(\frac{y_R - y_L}{x_R - x_L}\right) and interocular distance Deyes=(xR−xL)2+(yR−yL)2D_{\text{eyes}} = \sqrt{(x_R - x_L)^2 + (y_R - y_L)^2}
    Rotate image II by −ϕ-\phi to set eye-to-eye axis horizontal
    Rescale image such that DeyesD_{\text{eyes}} matches a fixed reference interocular distance
    return InormI_{\text{norm}}

  5. Knowl 5 — Discrimination Ranking of Facial Feature Templates

    empirical result

    When automated face recognition is performed using individual facial sub-templates evaluated separately on frontal face images, the discriminatory power of the individual facial features ranks in the following strictly decreasing order:

    1. Eyes: Yields the highest individual recognition accuracy and the highest class separation margin.
    2. Nose: Intermediate individual performance, slightly lower than the eyes.
    3. Mouth: Lower recognition accuracy than both the eyes and nose.
    4. Whole Face Template: Exhibits the lowest individual recognition performance among all tested templates.

    The whole face template performs worst because:

    • Achieving precise global scale and orientation normalization across the entire head outline is difficult.
    • Whole-face correlation is highly sensitive to slight 3D head rotations (out-of-plane deviations from strict frontal views) and variations in hair style, whereas smaller localized sub-templates (eyes, nose, mouth) behave as approximately planar patches and remain distortion-resistant.

  6. Knowl 6 — Bayes Nearest-Neighbor Classifier with Pooled Covariance for Facial Feature Vectors

    model/method

    To classify an unlabelled 35-dimensional geometrical facial feature vector x∈R35x \in \mathbb{R}^{35}, the feature distribution for each individual subject j∈{1,…,N}j \in \{1, \dots, N\} is modeled as a multivariate Gaussian distribution N(mj,Σ)\mathcal{N}(m_j, \Sigma). Each subject jj has a class-specific mean vector mjm_j, while all classes share a common pooled covariance matrix Σ\Sigma:

    Σ=1N∑i=1NΣi\Sigma = \frac{1}{N} \sum_{i=1}^N \Sigma_i

    where Σi\Sigma_i is the empirical covariance matrix estimated from the available training samples of subject ii.

    The distance from feature vector xx to subject class jj is the Mahalanobis distance:

    dj(x)=(x−mj)TΣ−1(x−mj)d_j(x) = (x - m_j)^T \Sigma^{-1} (x - m_j)

    The classifier assigns xx to the subject index j∗j^* with minimum distance:

    j∗=arg⁡min⁡j∈{1,…,N}dj(x)j^* = \arg\min_{j \in \{1, \dots, N\}} d_j(x)

    To allow classification rejection, a distance threshold τ\tau is introduced: if min⁡jdj(x)>τ\min_j d_j(x) > \tau, classification is suspended (the probe is rejected as unrecognized).

  7. Knowl 7 — Effect of Image Preprocessing on Template Correlation Under Illumination Variation

    data/table

    The robustness of normalized cross-correlation template matching against illumination changes was evaluated across images of the same subject under different lighting angles and shadowing. The dependency of the matching correlation score on lighting changes is measured by the correlation reduction ratio: the ratio of the average cross-correlation between identical faces under varying illumination to the cross-correlation between matching reference faces under controlled illumination. A lower ratio indicates greater invariance to lighting changes.

    PreProcessing Method Corr. Reduction Ratio
    No Normalization (II) 1.50
    Intensity Ratio (I/⟨I⟩I / \langle I \rangle) 1.20
    Gradient Magnitude (∣∂xI∣+∣∂yI∣|\partial_x I| + |\partial_y I|) 1.18
    Laplacian (∣∂xxI+∂yyI∣|\partial_{xx} I + \partial_{yy} I|) 1.19

    Precomputing the gradient magnitude D(I)=∣∂xI∣+∣∂yI∣D(I) = |\partial_x I| + |\partial_y I| (calculated using an L1L_1 norm on Gaussian-regularized image intensities) achieves the lowest correlation reduction ratio (1.181.18) and yields the highest invariance against illumination variations and shadowing.

  8. Knowl 8 — Chin Contour Detection via Dynamic Programming in Elliptical Coordinates

    algorithm

    To extract the chin contour from a normalized frontal face image, the image is mapped into an elliptical coordinate system to exploit the natural elliptical shape of the human lower face, and dynamic programming is used to locate the optimal edge path:

    Input: Grayscale face image I(x,y)I(x,y) with detected eye centers and estimated nose tip position.
    Output: Vector of 11 radial distances describing the chin contour.
    Define an elliptical coordinate frame (ρ,θ)(\rho, \theta) centered at the mid-face with major/minor axes aligned with expected facial proportions
    Compute vertical/radial gradient intensity map G(ρ,θ)=∣∇I(ρ,θ)∣G(\rho, \theta) = |\nabla I(\rho, \theta)| within an elliptical search annulus bounded around the expected chin location
    Formulate the chin boundary extraction as finding a path ρ(θ)\rho(\theta) that maximizes edge intensity while penalizing excessive deviation from an elliptical contour:
        Cost Function: E(ρ(θ))=∑θ[−G(ρ(θ),θ)+λ⋅(ρ(θ)−ρ0)2+μ⋅(dρdθ)2]E(\rho(\theta)) = \sum_{\theta} \left[ -G(\rho(\theta), \theta) + \lambda \cdot (\rho(\theta) - \rho_0)^2 + \mu \cdot \left(\frac{d\rho}{d\theta}\right)^2 \right]
        where ρ0\rho_0 is the expected elliptical radius, and λ,μ\lambda, \mu are weighting parameters
    Solve for the optimal contour ρ∗(θ)\rho^*(\theta) via 1D serial dynamic programming across discretized angular intervals θ∈[θmin⁡,θmax⁡]\theta \in [\theta_{\min}, \theta_{\max}]
    Sample ρ∗(θ)\rho^*(\theta) at 11 equiangular radii originating from the mouth center downwards to the chin boundary
    return 11-dimensional chin radius feature vector

  9. Knowl 9 — Multiresolution Robustness of Template Matching Face Recognition

    empirical result

    When template matching is evaluated across a 4-level Gaussian multiresolution pyramid (downsampling by factors of 1÷81 \div 8 relative to the original 512×512512 \times 512 image):

    • Correct recognition performance remains stable and robust from full resolution down to a 1:41:4 subsampling ratio, corresponding to an interocular distance of approximately 27–3027\text{--}30 pixels.
    • At 1:41:4 subsampling, facial sub-templates are as small as 36×3636 \times 36 pixels, yet recognition accuracy and class separation remain near peak levels.
    • Significant performance degradation occurs only when resolution is degraded beyond 1:41:4 (e.g., at 1:81:8 subsampling where interocular distance drops below 15 pixels).

    This demonstrates that template matching for face recognition does not require high-resolution imagery and can operate effectively with low computational cost at coarse image scales.

  10. Knowl 10 — Min/Max Ratio Metric for Face Classification Robustness

    definition

    The Min/Max ratio (RmMR_{mM}) is a scalar performance metric quantifying the separation margin and classification robustness of a nearest-neighbor classifier across a multi-class dataset.

    For a dataset of NN classes where each class j∈{1,…,N}j \in \{1, \dots, N\} is represented by a prototype mjm_j in feature space (or a reference template), the Min/Max ratio for class jj is defined as:

    RmM(j)=min⁡k≠jd(mj,mk)max⁡x∈Cjd(x,mj)R_{mM}(j) = \frac{\min_{k \ne j} d(m_j, m_k)}{\max_{x \in C_j} d(x, m_j)}

    where:

    • min⁡k≠jd(mj,mk)\min_{k \ne j} d(m_j, m_k) is the minimum distance from class jj's prototype to any incorrect class prototype (minimum inter-class separation).
    • max⁡x∈Cjd(x,mj)\max_{x \in C_j} d(x, m_j) is the maximum distance between prototype mjm_j and any training/testing sample xx belonging to class CjC_j (maximum intra-class dispersion).

    The overall dataset metric is the average ratio RmM=1N∑j=1NRmM(j)R_{mM} = \frac{1}{N} \sum_{j=1}^N R_{mM}(j). An RmM>1R_{mM} > 1 for all classes guarantees zero classification errors (100%100\% correct classification) under nearest-neighbor decision rules.

Coverage note — None. All major contributions—including the geometrical feature extraction pipeline, the multi-region template matching system, hierarchical normalization, classification schemes, and comparative empirical evaluations—have been converted into knowls.

References

  1. 1.R. J. Baron. Mechanisms of human facial recognition. International Journal of Man-Machine Studies, 15:137–178, 1981.
  2. 2.Martin Bichsel. Strategies of Robust Object Recognition for the Identification of Human Faces. PhD thesis, Eidgenossischen Technischen Hochschule, Zurich, 1991.
  3. 3.W. W. Bledsoe. Man-machine facial recognition. Technical Report Rep. PRI:22, Panoramic Research Inc., Paolo Alto, Cal., 1966.
  4. 4.R. Brunelli. Edge projections for facial feature extraction. Technical Report 9009-12, I.R.S.T., 1990.
  5. 5.R. Brunelli. Face recognition: Dynamic programming for the detection of face outline. Technical Report 9104-06, I.R.S.T., 1991.
  6. 6.R. Brunelli and T. Poggio. Hyperbf networks for real object recognition. In Morgan-Kauffman, editor, Proc. 12th IJCAI, Sidney, 1991.
  7. 7.R. Brunelli and T. Poggio. Hyperbf Networks for Gender Classification. In Proc. DARPA Image Understanding Workshop, 1992.
  8. 8.J. Buhmann, J. Lange, and C. von der Malsburg. Distortion invariant object recognition by matching hierarchically labeled graphs. In Proceedings of IJCNN'89, pages 151–159, 1989.
  9. 9.D. J. Burr. Elastic matching of line drawings. IEEE Transactions on Pattern Analysis and Machine Intelligence, 3(6):708–713, 1981.
  10. 10.P. J. Burt. Smart sensing within a pyramid vision machine. Proceedings of the IEEE, 76(8):1006–1015, 1988.
  11. 11.H. Chan and W. W. Bledsoe. A man-machine facial recognition system: some preliminary results. Technical report, Panoramic Research Inc., Cal, 1965.
  12. 12.G. Cottrell and M. Fleming. Face recognition using unsupervised feature extraction. In Proceedings of the International Neural Network Conference, 1990.
  13. 13.I. Craw, H. Ellis, and J.R. Lishman. Automatic extraction of face features. Pattern Recognition Letters, 5:183–187, Feb. 87.
  14. 14.K. Fukunaga. Introduction to Statistical Pattern Recognition. Academic Press, 1972.
  15. 15.A. J. Goldstein, L. D. Harmon, and A. B. Lesk. Identification of human faces. In Proc. IEEE, Vol. 59, page 748, 1971.
  16. 16.B. A. Golomb, D. T. Lawrence, and T. J. Sejnowski. Sexnet: A neural network identifies sex from human faces. In Advances in Neural Information Processing Systems 3, pages 572–577, 1991.
  17. 17.Zi-Quan Hong. Algebraic feature extraction of image for recognition. Pattern Recognition, 24(3):211–219, 1991.
  18. 18.T. Kanade. Picture processing by computer complex and recognition of human faces. Technical report, Kyoto University, Dept. of Information Science, 1973.
  19. 19.Y. Kaya and K. Kobayashi. A basic study on human face recognition. In S. Watanabe, editor, Frontiers of Pattern Recognition, page 265. 1972.
  20. 20.M. Kirby and L. Sirovich. Application of the Karhunen-Loeve procedure for the characterization of human faces. IEEE Transactions on Pattern Analysis and Machine Intelligence, 12(1):103–108, 1990.
  21. 21.Y. Lee. Handwritten digit recognition using k nearest-neighbor, radial basis functions and backpropagation neural networks. Neural Computation, 3(3), 1991.
  22. 22.T. Matsuhashi, S. Matusi, and T. Minami. Identification of human faces based on isodensity maps. Pattern Recognition, 24(3):263–272, 1991.
  23. 23.T. Poggio. A Theory of How the Brain Might Work, volume LV of Cold Spring Harbor Symposia on Quantitative Biology, pages 899–910. Cold Spring Harbor Laboratory Press, 1990.
  24. 24.T. Poggio. 3d object recognition and prototypes: One 2d view may be sufficient. Technical Report 9107-02, I.R.S.T., 1991.
  25. 25.T. Poggio and S. Edelman. A Network that Learns to Recognize Three-Dimensional Objects. Nature, 343(6225):1–3, 1990.
  26. 26.T. Poggio and F. Girosi. A theory of networks for approximation and learning. A.I. Memo No. 1140, Massachusetts Institute of Technology, 1989.
  27. 27.T. Poggio and F. Girosi. Networks for Approximation and Learning. In Proc. of the IEEE, Vol. 78, pages 1481–1497, 1990.
  28. 28.T. Poggio and L. Stringa. A Project for an Intelligent System: Vision and Learning. Internation Journal of Quantum Chemistry, 42:727–739, 1992.
  29. 29.J. Sergent. Structural processing of faces. In A.W. Young and H.D. Ellis, editors, Handbook of Research on Face Processing. North-Holland, Amsterdam, 1989.
  30. 30.M. Dalla Serra and R. Brunelli. On the Use of the Karhunen-Loeve Expansion for Face Recognition. Technical Report 9206-04, I.R.S.T., 1992.
  31. 31.Luigi Stringa. Automatic Face Recognition using Directional Derivatives. Technical Report 9205-04, I.R.S.T., 1991.
  32. 32.Luigi Stringa. Eyes Detection for Face Recognition. Technical Report 9203-07, I.R.S.T., 1991.
  33. 33.M. Turk and A. Pentland. Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3(1):71–86, 1991.
  34. 34.H. Voorhees. Finding texture boundaries in images. Technical Report AI-TR 968, M.I.T. Artificial Intelligence Laboratory, 1987.
  35. 35.A. W. Young and H. D. Ellis, editors. Handbook of Research on Face Processing. NORTH-HOLLAND, 1989.
  36. 36.Alan L. Yuille. Deformable templates for face recognition. Journal of Cognitive Neuroscience, 3(1):59–70, 1991.

Citation

MLA
Brunelli, R., and T. Poggio. “Face Recognition: Features Versus Templates”. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 15, no. 10, 1993, pp. 1042–52, https://doi.org/10.1109/34.254061.
APA
Brunelli, R., & Poggio, T. (1993). Face recognition: features versus templates. IEEE Transactions on Pattern Analysis and Machine Intelligence, 15(10), 1042–1052. https://doi.org/10.1109/34.254061
Chicago
Brunelli, R., and T. Poggio. 1993. “Face Recognition: Features Versus Templates”. IEEE Transactions on Pattern Analysis and Machine Intelligence 15 (10): 1042–52. https://doi.org/10.1109/34.254061.
Harvard
Brunelli, R. and Poggio, T. (1993) “Face recognition: features versus templates”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 15(10), pp. 1042–1052. Available at: https://doi.org/10.1109/34.254061.
Vancouver
1. Brunelli R, Poggio T (1993) Face recognition: features versus templates. IEEE Transactions on Pattern Analysis and Machine Intelligence 15:1042–1052

BibTeX

@article{Brunelli_1993, title={Face recognition: features versus templates}, volume={15}, ISSN={0162-8828}, url={http://dx.doi.org/10.1109/34.254061}, DOI={10.1109/34.254061}, number={10}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, publisher={Institute of Electrical and Electronics Engineers (IEEE)}, author={Brunelli, R. and Poggio, T.}, year={1993}, pages={1042–1052} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF