Back

Face Recognition: The Problem of Compensating for Changes in Illumination Direction

Yael AdiniY. MosesS. Ullman

article1994TPAMI1,443 citations

Demonstrates through systematic empirical testing that standard illumination-invariant image representations such as edge maps, intensity derivatives, and 2D Gabor filters fail to overcome lighting direction changes in face recognition, establishing the need for richer three-dimensional or model-based recognition strategies.

Listen

Automated face recognition systems face a major practical hurdle when identifying individuals across varying real-world conditions, particularly when lighting directions shift. In practical security, surveillance, and verification environments, illumination differences can alter an image more drastically than the physical differences between two separate individuals. To address this, many existing computer vision approaches rely on early-stage, general image representations—such as edge maps, image derivatives, and spatial frequency filters—assuming these representations are largely insensitive to lighting changes.

The article systematically evaluates whether these widely used early-stage image filtering representations and standard distance comparison measures are sufficient on their own to recognize faces across changes in illumination direction, viewpoint, and facial expression.

To test this question, the authors conducted an empirical study using a tightly controlled dataset of 25 individuals without distinct features like facial hair or glasses. The database isolated specific parameters, including left versus right illumination, frontal versus rotated viewpoints (34 degrees), and neutral versus expressive faces. The authors evaluated 107 distinct operator parameter combinations across several core representations, including raw gray levels, edge maps, directional and nondirectional Gaussian derivatives, and two-dimensional Gabor-like filters, alongside logarithmic intensity transformations. These transformed images were evaluated across three facial regions using five standard mathematical distance metrics, generating approximately 100,000 pairwise comparisons.

The findings demonstrate that none of the evaluated representations are sufficient by themselves to overcome lighting direction changes. Using unprocessed gray-level images resulted in a complete failure, yielding a 100 percent miss rate where lighting variations completely masked individual identity. Across all 107 processed representations, the majority exhibited miss rates exceeding 50 percent. Representations tuned to horizontal features performed best—likely because natural facial features such as the eyes and mouth run horizontally and are less disrupted by horizontal light shifts—yet even the most optimal configuration failed to recognize 20 percent of the faces in the database and produced an average failure rate above 40 percent. Similar severe recognition failures occurred when viewing angles rotated by 34 degrees or when faces adopted extreme expressions, whereas human observers given the same dataset achieved identification accuracy exceeding 97 percent.

These results demonstrate that early-stage visual filtering and universal edge representations cannot compensate for lighting and viewpoint variations on their own. System designs that rely strictly on these representations face serious performance risks, high error rates, and vulnerability to spoofing or misidentification under uncontrolled lighting. Furthermore, the stark performance gap between these algorithms and human visual perception indicates that robust recognition requires higher-level processing rather than just early sensory filtering.

Moving forward, developers should look beyond generic low-level filters and implement domain-specific methods. Supported alternatives include model-based approaches that leverage three-dimensional geometry and surface reflectance, class-based techniques that exploit structural face properties such as bilateral symmetry and stable feature boundaries, or multi-image models that sample varied lighting conditions. When designing and evaluating face recognition systems, teams should conduct isolated parameter testing against rigorously controlled benchmarks before full-scale deployment.

The primary limitation of the study is its evaluation of individual, isolated parameter changes rather than complex simultaneous combinations of scale, background, and multi-directional lighting shifts. However, because simple representations failed even under these controlled, single-variable tests, there is high confidence in the conclusion that early filtering alone cannot solve the illumination problem in face recognition.

Weizmann Institute of Scienceedge mapfacial expressionillumination conditionsface recognitionillumination directionillumination changes
Adini et al (1994).pdf
Cover for Face Recognition: The Problem of Compensating for Changes in Illumination Direction

Abstract

A face recognition system must recognize a face from a novel image despite the variations between images of the same face. A common approach to overcoming image variations because of changes in the illumination conditions is to use image representations that are relatively insensitive to these variations. Examples of such representations are edge maps, image intensity derivatives, and images convolved with 2D Gabor-like filters. Here we present an empirical study that evaluates the sensitivity of these representations to changes in illumination, as well as viewpoint and facial expression. Our findings indicated that none of the representations considered is sufficient by itself to overcome image variations because of a change in the direction of illumination. Similar results were obtained for changes due to viewpoint and expression. Image representations that emphasized the horizontal features were found to be less sensitive to changes in the direction of illumination. However, systems based only on such representations failed to recognize up to 20 percent of the faces in our database. Humans performed considerably better under the same conditions. We discuss possible reasons for this superiority and alternative methods for overcoming illumination effects in recognition.

Topics

edge mapfacial expressionillumination conditionsimage representationslighting conditionsviewpointface recognition systemsfacial images

Keywords

face recognitionillumination directionillumination changesimage representationfacial featuresgray-level imagesreflectance propertiesedge detectionhuman visual system

Table of Contents

  • 1 INTRODUCTION
  • 1.1 Edge Map
  • 1.2 The Image Filtered with 2D Gabor-Like Functions
  • 1.3 Derivatives of the Gray-Level
  • 1.4 Log Transformations
  • 2 METHODS
  • 3 RESULTS
  • 3.1 Illumination Direction
  • 3.2 Viewpoint
  • 3.3 Expression
  • 4 SUMMARY AND DISCUSSION
  • 4.1 Evaluating Recognition Systems
  • 4.2 Alternative Approaches
  • 4.2.1 Independent Image Comparisons
  • 4.2.2 The Model-Based Approaches
  • 4.2.3 The Class-Based Approaches
  • 4.3 Comparisons with the Human Visual System
  • APPENDIX: METHODS
  • A.1 The Database of Faces
  • A.2 Distance Measures
  • A.3 Image Representations
  • A.3.1 Edge Representation
  • A.3.2 Laplacian-of-Gaussian Filter
  • A.3.3 Gray-Level Derivatives
  • A.3.4 2D Gabor-Like Filters
  • A.3.5 Log Representation
  • ACKNOWLEDGMENTS
  • REFERENCES

Knowls

  1. Knowl 1 — Empirical Insufficiency of Early Image Representations for Compensating Illumination Direction

    empirical result

    Evaluating 107 image representations (including raw gray levels, edge maps from the Shen-Castan DRF detector, Laplacian-of-Gaussian filters, first- and second-order directional Gaussian derivatives, 2D Gabor-like filters, and logarithmic intensity transformations) paired with five distance measures on a database of 25 normalized male faces revealed that universal, domain-independent representations cannot eliminate the variations caused by changes in illumination direction:

    • When comparing raw, unprocessed gray-level images under left vs. right directional illumination, a nearest-neighbor classification scheme suffered a 100%100\% miss-percentage and a 100%100\% failure-rate. This confirms that image variations induced by shifting the light source across the horizontal axis are strictly larger than the inter-individual differences between different people under identical illumination.
    • Applying low-level filtering representations improved performance over raw intensities but failed to overcome the illumination variations. Across all 107 operators, miss-percentages ranged from 20%20\% to 100%100\%, with the majority exceeding 50%50\%.
    • Even under the single best combination of operator, distance metric, and face mask, the system failed to recognize 20%20\% of the individuals (5 out of 25 faces), with an associated failure-rate of approximately 40%40\% on those missed faces (confusing each missed individual with multiple non-matching database faces).

  2. Knowl 2 — Illumination Robustness Advantage of Horizontally Tuned Image Filters

    empirical result

    Among all spatial filtering representations evaluated for compensating horizontal illumination changes (left vs. right light source), filter orientation has the strongest effect on recognition accuracy:

    • 2D Gabor-like filters and directional Gaussian derivatives tuned to horizontal image features (orientation θ=90∘\theta = 90^\circ relative to the vertical axis, or vertical spatial derivative dyd_y) achieved the lowest miss-percentage of 20%20\% at optimal spatial scales.
    • In contrast, identical filter formulations oriented vertically (θ=0∘\theta = 0^\circ) or diagonally (θ=45∘,135∘\theta = 45^\circ, 135^\circ) produced miss-percentages exceeding 60%60\%.
    • This orientation dependence arises because key facial landmarks (eyebrows, eyes, lips) are horizontally oriented, making their albedo boundaries relatively stable across lighting shifts, whereas horizontal displacements of the light source produce dominant vertical intensity gradients and cast shadows that severely degrade vertically oriented filters.

  3. Knowl 3 — Distance Measures for Comparing Gray-Level Image Representations

    equation

    Let I1(x)I_1(x) and I2(x)I_2(x) denote the scalar values of pixel location xx in two registered image representations, and let mask\text{mask} denote the set of nn valid pixels comprising a defined facial region. Dissimilarity between face representations is evaluated using five distance functions:

    1. Pointwise Distance (L1L_1-norm average pixel difference): Pointwise(I1,I2)=1n∑x∈mask∣I1(x)−I2(x)∣\text{Pointwise}(I_1, I_2) = \frac{1}{n} \sum_{x \in \text{mask}} |I_1(x) - I_2(x)|

    2. Regional Distance (displacement-tolerant difference within a 5×55 \times 5 pixel neighborhood neighb(x)\text{neighb}(x)): Regional(I1,I2)=1n∑x∈maskmin⁡i∈neighb(x)∣I1(x)−I2(i)∣\text{Regional}(I_1, I_2) = \frac{1}{n} \sum_{x \in \text{mask}} \min_{i \in \text{neighb}(x)} |I_1(x) - I_2(i)| which tolerates spatial misalignments of up to ±2\pm 2 to 33 pixels in the image plane.

    3. Affine-GL Distance (uniform affine illumination normalization): Affine-GL(I1,I2)=min⁡a,b∈R1n∑x∈mask(aI1(x)+b−I2(x))2\text{Affine-GL}(I_1, I_2) = \min_{a, b \in \mathbb{R}} \frac{1}{n} \sum_{x \in \text{mask}} \left(a I_1(x) + b - I_2(x)\right)^2 where the scalar gain aa and offset bb are computed analytically via linear least squares.

    4. Local Affine-GL Distance: Computed by partitioning the face mask into disjoint square blocks of 16×1616 \times 16 pixels, calculating the optimal affine scalar transformation independently within each block, and averaging the resulting squared errors across all blocks.

    5. LOG Distance: The pointwise distance evaluated on the base-2 logarithm of the pixel intensities: LOG(I1,I2)=1n∑x∈mask∣log⁡2(I1(x))−log⁡2(I2(x))∣\text{LOG}(I_1, I_2) = \frac{1}{n} \sum_{x \in \text{mask}} |\log_2(I_1(x)) - \log_2(I_2(x))|

    Across the 107 representations tested, Local Affine-GL and Pointwise distance yielded the lowest miss-percentages.

  4. Knowl 4 — Error Metrics for Illumination Sensitivity: Missed-Face, Miss-Percentage, and Failure-Rate

    definition

    To evaluate how well an image representation and distance metric D(IA,IB)D(I_A, I_B) distinguish images of the same individual under different viewing conditions from images of different individuals under identical viewing conditions, three metrics are defined:

    • Missed-Face: A face identity kk is defined as a missed-face if the distance between two images of person kk taken under different imaging conditions (e.g., left vs. right illumination) is larger than the distance between person kk's image and an image of another person j≠kj \neq k taken under the reference condition: D(Ikcondition 1,Ikcondition 2)>D(Ikcondition 1,Ijcondition 1)D(I_k^{\text{condition 1}}, I_k^{\text{condition 2}}) > D(I_k^{\text{condition 1}}, I_j^{\text{condition 1}})

    • Miss-Percentage: The percentage of face identities in the database classified as missed-faces: miss-percentage=number of missed facestotal number of face identities×100%\text{miss-percentage} = \frac{\text{number of missed faces}}{\text{total number of face identities}} \times 100\% where 0%0\% represents perfect recognition and 100%100\% denotes complete failure.

    • Failure-Rate: The average percentage of non-matching individuals in the database whose distance to the target image is smaller than the distance between the matching pair under the altered condition. A high failure-rate demonstrates that recognition errors reflect widespread confusion across the database rather than isolated, accidental similarities between specific pairs.

  5. Knowl 5 — 2D Gabor-like Filter Representations for Face Images

    model/method

    To model biological receptive fields in early visual cortex (simple cells in V1) and extract localized orientation and spatial frequency components, 2D Gabor-like filters with unity aspect ratio (σy/σx=1\sigma_y / \sigma_x = 1) are applied to face images. The even-symmetric (cosine) and odd-symmetric (sine) filter kernels are defined as:

    CosG(x,y)=cos⁡(2πλ(xcos⁡θ+ysin⁡θ))exp⁡(−x2+y22σ2)\text{CosG}(x, y) = \cos\left(\frac{2\pi}{\lambda}(x \cos\theta + y \sin\theta)\right) \exp\left(-\frac{x^2 + y^2}{2\sigma^2}\right)

    SinG(x,y)=sin⁡(2πλ(xcos⁡θ+ysin⁡θ))exp⁡(−x2+y22σ2)\text{SinG}(x, y) = \sin\left(\frac{2\pi}{\lambda}(x \cos\theta + y \sin\theta)\right) \exp\left(-\frac{x^2 + y^2}{2\sigma^2}\right)

    where:

    • θ∈{0∘,45∘,90∘,135∘}\theta \in \{0^\circ, 45^\circ, 90^\circ, 135^\circ\} denotes the harmonic modulation orientation relative to the vertical axis (θ=90∘\theta = 90^\circ corresponds to horizontal selectivity).
    • λ\lambda is the spatial wavelength of the harmonic modulation in pixels.
    • σ\sigma is the standard deviation of the Gaussian envelope, parameterized such that σ=λ/2\sigma = \lambda / 2, with a spatial mask support size of 2λ2\lambda.
    • Evaluated scales include σ∈{2,4,6,8,10,12,14,16,18,20}\sigma \in \{2, 4, 6, 8, 10, 12, 14, 16, 18, 20\} pixels, corresponding to λ∈[4,40]\lambda \in [4, 40] pixels (roughly 4 to 40 cycles per face height in a 170×170170 \times 170 pixel face image).

    The filtered representation is computed by 2D spatial convolution of the normalized grayscale face image with the specified Gabor kernel.

  6. Knowl 6 — Parametric Effects of Filter Scale, Filter Symmetry, and Face Region Masking

    empirical result

    Systematic evaluation of parametric settings for spatial filter representations across the 25-face database established the following effects:

    • Filter Scale: Spatial scale strongly modulates illumination sensitivity. For 2D Gabor filters, intermediate wavelengths λ∈[12,24]\lambda \in [12, 24] pixels (Gaussian standard deviation σ∈[6,12]\sigma \in [6, 12] pixels, corresponding to approximately 7 to 14 cycles per face diameter) achieved the lowest miss-percentage (20%20\%). At very large scales (λ=32,σ=16\lambda = 32, \sigma = 16), despite severe visual blurring where individual facial features became unrecognizable, the miss-percentage remained comparatively low (36%36\%).
    • Filter Symmetry: Even-symmetric (cosine) and odd-symmetric (sine) Gabor filters showed virtually indistinguishable miss-percentages across all scales and orientations, indicating that phase symmetry does not affect illumination compensation.
    • Face Region Masking: Evaluating discrimination across three facial subregions (full face excluding hair, eyes region, and lower face/mouth) revealed that the inner features mask (eyes region) yielded the lowest miss-percentages, whereas the lower facial region yielded the highest miss-percentages due to high sensitivity to illumination variations and lower distinctiveness.

  7. Knowl 7 — Sensitivity of Low-Level Image Representations to Viewpoint and Expression Variations

    empirical result

    Evaluating low-level image representations against non-illumination image transformations revealed significant limitations:

    • Viewpoint Changes (34∘34^\circ horizontal head rotation): Comparing raw gray-level images resulted in a 100%100\% miss-percentage. Across all 107 filtered representations, none compensated adequately for a 34∘34^\circ horizontal viewpoint shift, with all operators exhibiting miss-percentages strictly exceeding 50%50\%.
    • Mild Expression Changes (Smile): Direct pixelwise comparison of raw, unfiltered images using the full-face mask achieved a 0%0\% miss-percentage (perfect identification). However, applying spatial filters designed for illumination invariance (such as odd-symmetric horizontal Gabor filters) degraded performance, increasing the smile miss-percentage to 34%34\%.
    • Drastic Expression Changes (Closed eyes and wide-open mouth): Comparing raw gray-level images yielded a 60%60\% miss-percentage across the full face and >80%>80\% in the eye region. Applying low-level filtering operators failed to improve these results and often increased the miss-percentage, demonstrating that early filtering stages designed for one invariant can interfere with tolerance to geometric shape changes.

  8. Knowl 8 — Human Generalization Superiority and Face Inversion Effects under Lighting Shifts

    empirical result

    Psychophysical experiments using the identical controlled face database established a clear performance gap between human perception and universal computational representations:

    • Human observers trained to identify upright faces achieved >97%>97\% identification accuracy when tested on novel images of those individuals under altered illumination (left vs. right) and viewpoint (34∘34^\circ yaw).
    • When faces were presented inverted (upside down) during both training and testing, human generalization accuracy dropped to 89%89\% for novel illumination directions and 85%85\% for novel viewpoints.
    • Because low-level spatial filtering and edge detection mechanisms operate identically on upright and inverted images, the drop in human generalization performance under inversion indicates that human illumination compensation does not rely solely on early visual filtering (such as in V1), but requires higher-level, orientation-dependent, class- or object-specific mechanisms located in higher visual cortical areas (such as the inferior temporal cortex).

  9. Knowl 9 — Standardized Acquisition and Geometric Normalization Protocol for the Face Database

    experimental setup

    The empirical study utilized a standardized subset of the Weizmann Facebase consisting of 25 male subjects without distinctive extraneous features (no glasses, beards, or mustaches).

    Acquisition Setup:

    • Images were captured using a Pulnix TM-560 camera with a Canon V6x16 16–100mm F1:1.9 lens mounted on an Adept One robotic arm at a fixed distance of 110 cm against a uniform wooden background.
    • Prior to image capture, each subject's face was geometrically normalized using optical reference lines aligning the vertical facial symmetry axis, the external corners of both eyes, and the base of the nose.
    • Initial images (512×352512 \times 352 pixels, 8 bits/pixel) were subsampled by a factor of 2 to 256×176256 \times 176 pixels, with the face region occupying approximately 170×170170 \times 170 pixels.

    Evaluated Image Conditions per Subject:

    1. Frontal viewpoint, left directional illumination, neutral expression.
    2. Frontal viewpoint, right directional illumination, neutral expression.
    3. 34∘34^\circ left horizontal viewpoint rotation, left illumination, neutral expression.
    4. Frontal viewpoint, left illumination, smiling expression.
    5. Frontal viewpoint, left illumination, drastic expression (closed eyes, open mouth).

  10. Knowl 10 — Mathematical Formulations of Differential and Edge Operators for Illumination Compensation

    model/method

    In addition to 2D Gabor filters, three classes of differential and edge operators were implemented and tested:

    1. Laplacian-of-Gaussian Filter (LoG): ∇2G(r)=−1πσ4(1−r22σ2)exp⁡(−r22σ2)\nabla^2 G(r) = \frac{-1}{\pi \sigma^4} \left(1 - \frac{r^2}{2\sigma^2}\right) \exp\left(-\frac{r^2}{2\sigma^2}\right) where r2=x2+y2r^2 = x^2 + y^2, evaluated at Gaussian smoothing scales σ∈{2,4}\sigma \in \{2, 4\} pixels.

    2. Smoothed Gray-Level Derivatives: Formed by convolving the image with first-order spatial derivatives of an isotropic Gaussian at standard deviations σ∈{6,8,12,16,20}\sigma \in \{6, 8, 12, 16, 20\} pixels:

    • Isotropic radial derivative: dr=−rσ2exp⁡(−r22σ2)d_r = -\frac{r}{\sigma^2} \exp\left(-\frac{r^2}{2\sigma^2}\right)
    • Horizontal derivative: dx=−xσ2exp⁡(−x22σ2)d_x = -\frac{x}{\sigma^2} \exp\left(-\frac{x^2}{2\sigma^2}\right)
    • Vertical derivative: dy=−yσ2exp⁡(−y22σ2)d_y = -\frac{y}{\sigma^2} \exp\left(-\frac{y^2}{2\sigma^2}\right)
    1. Gaussian-Smoothed Edge Maps: Binary edge maps extracted using the Shen-Castan Difference of Recursive Filters (DRF) edge detector were converted into continuous gray-level representations by convolving with a 2D Gaussian filter of scale 2σ∈{5,11}\sqrt{2}\sigma \in \{5, 11\} pixels to allow distance metric evaluation.

Coverage note — None. The complete suite of mathematical formulations, experimental methodologies, parametric evaluations, and psychophysical comparisons contributed by the paper has been captured.

References

  1. 1.C. Blakemore and F. Campbell, “On the Existence of Neuron in the Human Visual System Selectively Sensitive to the Orientation and Size of Retinal Images,” J. Physiology, vol. 203, pp. 237-260, 1969.
  2. 2.B. Bruce, “Changing Faces: Visual and Non Visual Coding Processes in Face Recognition,” British J. Psychology, vol. 73, pp. 105-116, 1982.
  3. 3.R. Brunelli and T. Poggio, “HyperBF Networks for Real Object Recognition,” Proc. IJCAI, pp. 1,278-1,284, Sydney, Australia, 1991.
  4. 4.J. Buhmann, M. Lades, and F. Eeckman, “A Silicon Retina for Face Recognition,” Technical Report No. 8596-CS, Institut für Informatik, Universität Bonn, 1993.
  5. 5.J. burns, R. Weiss, and E. Riseman, “The Non-Existence of General-Case View-Variants,” Geometrical Invariance in Computer Vision, J. Mundy and A. Zisserman, eds. MIT Press, 1992.
  6. 6.F. Campbell and J. Robson, “Application of Fourier Analysis to the Visibility of Gratings,” J. Physiology, vol. 197, pp. 551-556, 1968.
  7. 7.J.F. Canny, “A Computational Approach to Edge Detection,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 8, pp. 679-698, 1986.
  8. 8.D. Clemens and D. Jacobs, “Space and Time Bounds on Indexing 3D Models from 2D Images,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 13, no. 10, pp. 1,007, 1,017, Oct. 1991.
  9. 9.J.G. Daugman, “Spatial Visual Channels in the Fourier Plane,” Vision Research, vol. 24, no. 9, pp. 891-910, 1984.
  10. 10.J.G. Daugman, “Uncertainty Relation for Resolution in Space, Spatial Frequency and Orientation, Optimized by Two Dimensional Cortical Filters,” J. Optical Soc. Am., vol. 2, pp. 1,160-1,169, 1985.
  11. 11.G.M. Davies, H. Ellis, and J.W. Shephered, “Face Recognition Accuracy as a Function of Mode of Representation,” J. Applied Psychology, vol. 92, pp. 507-523, 1978.
  12. 12.L.S. Davis, “A Survey of Edge Detection Techniques,” Computer Graphics and Image Processing, vol. 4, pp. 248-270, 1975.
  13. 13.R. De-Valois and K. De-Valois, Spatial Vision. New York: Oxford Univ. Press, 1990.
  14. 14.R. Deriche, “Optimal Edge Detection Using Recursive Filtering,” Proc. ICCV, pp. 501-504, London, 1987.
  15. 15.S. Edelman, D. Reisfeld, and Y. Yeshurun, “A System for Face Recognition that Learns from Examples,” Proc. European Conf. Computer Vision, S. Sandini, ed., pp. 787-791. Springer-Verlag, 1992.
  16. 16.A. Fioreniti, L. Maffei, and G. Sandini, “The Role of High Spatial Frequencies in Face Perception,” Perception, vol. 12, pp. 195-201, 1983.
  17. 17.V. Govindaraju, D.B. Sher, R. Srihari, and S.N. Srihari, “Locating Human Faces in Newspaper Photographs,” Proc. CVPR 89, pp. 549-554, 1989.
  18. 18.N. Graham and J. Nachmias, “Detection of Grating Pattern Containing Two Spatial Frequencies: A Composition of Single-Channel and Multi-Channel Models,” Vision Research, vol. 11, pp. 251-259, 1971.
  19. 19.P. Hallinan, “A Low-Dimensional Representation of Human Faces for Arbitrary Lighting Conditions,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, pp. 995-999, 1994.
  20. 20.R.M. Haralick, “Digital Step Edges from Zero Crossing of Second Directional Derivatives,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 6, pp. 58-68, 1984.
  21. 21.B.K.P. Horn and M. Brooks, Seeing Shape from Shading. Cambridge, Mass.: MIT Press, 1989.
  22. 22.T. Kanade, Computer Recognition of Human Faces. Basel and Stuttgart: Birkhauser Verlag, 1977.
  23. 23.J.J. Koenderink and A.J. van Doorn, “Receptive Field Families,” Biological Cybernetics, vol. 63, pp. 291-297, 1990.
  24. 24.B. Manjunath, C. Shekhar, R. Chellappa, and C. von der Malsburg, “A Robust Method for Detecting Image Features with Application to Face Recognition and Motion Correspondence,” ICPR B, pp. 208-212, 1992.
  25. 25.S. Marcelja, “Mathematical Description of the Responses of Simple Cortical Cells,” J. Optical Soc. Am., vol. 70, pp. 1,297-1,300, 1980.
  26. 26.D. Marr and E. Hildreth, “Theory of Edge Detection,” Proc. Royal Soc. London B, vol. 207, pp. 197-217, 1980.
  27. 27.Y. Moses, “Face Recognition: Generalization to Novel Images,” PhD thesis, Weizmann Inst. of Science, 1993.
  28. 28.Y. Moses, S. Edelman, and S. Ullman, “Generalization to Novel Images in Upright and Inverted Faces,” Technical Report cc93-14, The Weizmann Inst. of Science, 1993.
  29. 29.Y. Moses and S. Ullman, “Limitation of Non-Model-Based Recognition Schemes,” Proc. ECCV-92, G. Sandini, ed., pp. 820-828. Springer-Verlag, 1992.
  30. 30.A. O’Toole, R. Millward, and J. Anderson, “A Physical System Approach to Recognition Memory for Spatially Transformed Faces,” Neural Networks, vol. 1, pp. 179-199, 1988.
  31. 31.K. Patterson and A. Baddeley, “When Face Recognition Fails,” J. Experimental Psychology: Human Learning and Memory, vol. 3, pp. 406-417, 1977.
  32. 32.T. Poggio and S. Edelman, “A Network that Learns to Recognize Three Dimensional Objects,” Nature, vol. 343, pp. 263-266, 1990.
  33. 33.D. Pollen and S. Ronner, “Phase Relationships Between Adjacent Simple Cells in the Visual Cortex,” Science, pp. 1,409-1,411, 1981.
  34. 34.D. Pollen and S. Ronner, “Visual Cortical Neurons as Localized Spatial Frequency Filters,” IEEE Trans. Systems, Man, and Cybernetics, vol. 13, pp. 907-916, 1983.
  35. 35.D. Reisfeld and Y. Yeshurun, “Robust Detection of Facial Features by Generalized Symmetry,” ICPR A, pp. 117-120, 1992.
  36. 36.C.A. Rothwell, D.A. Forsyth, A. Zisserman, and J.L. Mundy, “Extracting Projective Structure from Single Perspective Views of 3D Point Sets,” Proc. ICCV-93, pp. 573-582, 1993.
  37. 37.J. Shen and S. Castan, “An Optimal Linear Operator for Edge Detection,” Proc. SPIE-87, 1987.
  38. 38.J. Shepherd, G. Davies, and H. Ellis, “Studies of Cue Saliency,” Perceiving and Recognizing Faces, G.M. Davies, H. Ellis, and J. Shepherd, eds., pp. 104-131. Academic Press, 1981.
  39. 39.C. Stromeyer and B. Julesz, “Spatial-Frequency Masking in Vision: Critical Bands and Spreads of Masking,” J. Optical Soc. Am., vol. 62, pp. 1,221-1,232, 1972.
  40. 40.V. Torre and T. Poggio, “On Edge Detection,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 8, pp. 147-163, 1986.
  41. 41.M. Turk and A. Pentland, “Eigenfaces for Recognition,” J. Cognitive Neuroscience, vol. 3, pp. 71-86, 1991.
  42. 42.S. Ullman and R. Basri, “Recognition by Linear Combinations of Models,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 13, pp. 992-1,005, 1991.
  43. 43.T. Vetter, T. Poggio, and H. Bulthoff, “The Importance of Symmetry and Virtual Views in Three Dimensional Object Recognition,” Current Biology, pp. 18-23, 1994.
  44. 44.E. Warrington and A. Taylor, “Two Categorical Stages of Object Recognition,” Perception, vol. 7, pp. 152-164, 1978.
  45. 45.L. Weiskrantz, “Visual Prototypes, Memory, and the Inferotemporal Lobe,” Vision, Memory, and the Temporal Lobe, E.I.M. Mishkin, ed., chapter 2, pp. 13-28. New York: Elsevier, 1990.
  46. 46.F. Werblin, “Control of Retinal Sensitivity,” J. General Physiology, vol. 63, pp. 62-87, 1974.
  47. 47.K. Wong, H. Law, and P. Tsang, “A System for Recognising Human Faces,” Proc. ICASSP, pp. 1,638-1,642, 1989.

Citation

MLA
Moses, Y., et al. “Face Recognition: The Problem of Compensating for Changes in Illumination Direction”. Lecture Notes in Computer Science, Springer Berlin Heidelberg, 1994, pp. 286–96, https://doi.org/10.1007/3-540-57956-7_33.
APA
Moses, Y., Adini, Y., & Ullman, S. (1994). Face recognition: The problem of compensating for changes in illumination direction. In Lecture Notes in Computer Science (pp. 286–296). Springer Berlin Heidelberg. https://doi.org/10.1007/3-540-57956-7_33
Chicago
Moses, Y., Y. Adini, and S. Ullman. 1994. “Face Recognition: The Problem of Compensating for Changes in Illumination Direction”. In Lecture Notes in Computer Science. Springer Berlin Heidelberg. https://doi.org/10.1007/3-540-57956-7_33.
Harvard
Moses, Y., Adini, Y. and Ullman, S. (1994) “Face recognition: The problem of compensating for changes in illumination direction”, Lecture Notes in Computer Science. Springer Berlin Heidelberg, pp. 286–296. Available at: https://doi.org/10.1007/3-540-57956-7_33.
Vancouver
1. Moses Y, Adini Y, Ullman S (1994) Face recognition: The problem of compensating for changes in illumination direction. In: Lecture Notes in Computer Science. Springer Berlin Heidelberg, pp 286–296

BibTeX

@inbook{Moses_1994, title={Face recognition: The problem of compensating for changes in illumination direction}, ISBN={9783540483984}, ISSN={1611-3349}, url={http://dx.doi.org/10.1007/3-540-57956-7_33}, DOI={10.1007/3-540-57956-7_33}, booktitle={Computer Vision — ECCV ’94}, publisher={Springer Berlin Heidelberg}, author={Moses, Yael and Adini, Yael and Ullman, Shimon}, year={1994}, pages={286–296} }
Metadata:Crossref

Access the Paper

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

Open PDF