Improving Shadow Suppression for Illumination Robust Face Recognition

• Face analysis methods that perform nearly perfectly under tightly controlled conditions face practical challenges in unconstrained scenes with varied expressions, occlusions, resolution, pose, and lighting.Source: §1 · p. 1
• Illumination varies with lighting intensity and direction, and its effects can exceed differences between individuals, making normalization important for face recognition.Source: §1 · p. 1
• The pipeline models face-image formation, generates a chromaticity-invariant representation, and uses shadow-edge detection to extend that invariance to color.It first separates highlighted and non-highlighted regions and uses Lambertian-surface and Planckian-lighting approximations.Source: §1 · pp. 1–2
• The method was evaluated on CMU-PIE and FRGC and improved state-of-the-art face-recognition performance, especially for hard shadows, qualitatively and quantitatively.Source: §1 · p. 2

• Holistic normalization adjusts grayscale intensity globally, but these methods do not model image formation and can struggle with soft shadows, hard shadows, and highlights.Histogram Equalization and Histogram Matching are examples; Gamma Intensity Correction maps image intensity.Source: §2 · p. 2
• Invariant-feature methods include frequency, Retinex-derived, local-normalization, and gradient-based approaches, but reported methods can lose edge detail or struggle with hard cast shadows and color-space extension.SQI is described as lacking edge preservation, while the Tan–Triggs chain and related approaches handle soft shadows but encounter hard-edged shadows.Source: §2 · pp. 2–3
• 3D-model methods use physical lighting models but face data-acquisition, computational-cost, and registration constraints; the proposed approach instead combines holistic normalization with a reflectance model without 3D priors.The paper also states that its color-space shadow removal can work jointly with grayscale lighting methods.Source: §2 · p. 3

3 SKIN COLOR ANALYSIS

• The section proposes a product-form face-skin color representation as the basis for its approach, using a model intended to account for diffuse color.Source: §3 · p. 3

3.1 Reflectance Model: Lambert vs. Phong

• Lambertian reflectance models diffuse skin while Phong adds a geometry- and lighting-dependent specular term; the method detects highlights first, then analyzes non-highlighted skin with the Lambertian model.The paper notes that skin is mostly Lambertian but highlights occur in practice due to oil coating and semi-transparent skin particles.Source: §3.1 · p. 3

3.2 Specular Highlight Detection

• Sparse specular coefficients motivate a two-component NMF with an L1 sparsity constraint, and the reported highlight detection is robust across lighting intensities and directions.The figure presents original images above detected highlight masks.Source: §3.2 · p. 4, Figure 3 · p. 4

3.3 Skin Color Formation

• The color-formation analysis models illumination as Planckian and skin as Lambertian, then uses sensor responses and spectral reflectance to derive RGB skin color.The paper treats the Planckian characteristic as approximately linear, permitting extension to bi- or multi-illuminant scenes; the derived representation motivates separating channel- and lighting-related components in chromaticity space.Source: §3.3 · p. 4

4.1 Skin Model in Chromaticity Space

• Geometric-mean-normalized chromaticity attenuates intensity, while the Lambertian skin assumption permits separating constant terms from channel- and lighting-related terms.The resulting representation is described as combining a lighting-invariant part with a lighting-related part.Source: §4.1 · p. 5

4.2 Chromaticity Invariant Image Generation

• Geometric-mean normalization places log-chromaticity points on a plane, and an orthogonal projection maps their 3D vectors to 2D coordinates.The plane follows from the normalized channels’ log values summing to zero.Source: §4.2 · p. 5
• Points from the same material form parallel lines in the projected plane; the nose-bridge line spans more lighting variation than the forehead line.The figure illustrates the original image, chromaticity points in 3D log space, and projected forehead and nose-bridge pixels.Source: §4.2 · p. 5, Figure 4 · p. 6

4.3 Entropy based Lighting Normalization

• At the entropy-minimizing angle, shadows are significantly suppressed in the chromaticity image; the projection direction is selected by minimizing projected-data entropy.The bin-width rule uses the interquartile range in place of standard deviation, making it more robust to outliers.Source: §4.3 · p. 6

4.4 Global Intensity Regularization

• Global intensity regularization estimates the dominant intensity with m = 0.1, scales its reference value to 0.5, and applies the same ratio to all pixels.The coefficient m reduces the impact of large values; its default follows Tan et al..Source: §4.4 · p. 6

5 SHADOW-FREE COLOR FACE RECOVERY

• The 1D chromaticity image loses texture and contrast, while reverse projection cannot reliably recover per-band intensity or preserve features such as mustaches and eyebrows.These shortcomings motivate additional analysis of RGB channels in log space and a full-color reconstruction module.Source: §5 · p. 7, §5.1 · p. 7

5.1 In-depth Analysis of 1D Chromaticity Image

• Because lighting changes slowly except at shadow edges, log-image spatial gradients are treated as invariant across lighting conditions away from those boundaries.The stated property applies to both x and y partial derivatives, including comparisons between illuminated and shadow regions.Source: §5.1 · p. 7

5.2 Shadow-Specific Edge Detection

• The shadow-specific edge detector maximizes entropy to choose a projection that preserves lighting variation, then compares projected chromaticity images to form an edge mask.Guided filtering is used to smooth the 2D chromaticity images before gradient calculation.Source: §5.2 · p. 7

5.3 Full Color Face Image Reconstruction

• The reconstruction forms a shadow-free gradient map for each log-RGB channel, derives a shadow-free Laplacian, and solves Poisson’s equation with zero Neumann boundary values and matrix-space discretization.A per-channel scale factor based on the brightest raw-image pixels is applied after exponentiation to recover color and intensity.Source: §5.3 · pp. 7–8

6.1 Databases and Experimental Settings

• Evaluation uses CMU-PIE frontal neutral-expression illumination views for identification and FRGC Exp.4 for verification, covering different lighting and pose conditions.The experiments select databases with varied lighting and full-color images.Source: §6.1 · p. 8

6.2 Visual Comparison and Discussion

• Visually, recovery removes hard and soft shadows, normalizes color under soft shading, and complements gray-scale methods against hard edges; soft-shadow edge ambiguity smooths facial detail.Hard shadow edges can be difficult for existing methods to distinguish from intrinsic texture.Source: §6.2 · pp. 9–11

6.3 Identification Results on CMU-PIE

• CII preprocessing generally improves CMU-PIE identification across gray-scale methods and features, and cast-shadow removal further improves VGG-Face performance.The reported exception is TT fusion, which fails to improve performance.Source: §6.3 · pp. 11–12, Table 2 · p. 11
• SQI/MSQ fusion can slightly hurt LBP and LPQ after edge smoothing, while LGBP alleviates this effect; TT fusion does not improve identification.The authors attribute the SQI/MSQ effect to sensitivity to smoothed shadow regions.Source: §6.3 · pp. 11–12

6.4 Verification Results on FRGC

• Recovered color preprocessing generally improves FRGC verification across nearly all gray-scale methods and features, with a relative 1.1% improvement for VGG-Face.The comparison reports verification rate at FAR = 0.1%.Source: Table 3 · p. 12, §6.4 · p. 13

7 CONCLUSION

• The approach improves face-recognition performance across illumination variations on different databases.The authors also describe its photo-realism and extensibility as promising for practical applications.Source: §7 · p. 13
• The pipeline introduces chromaticity space for face recognition, extracts intrinsic face images, and reconstructs full-color faces photo-realistically after shadow removal.The work explores physical interpretations of skin-color formation.Source: §7 · p. 13
• Future work will synthesize faces under different illumination conditions and combine pose-invariant techniques to address face analysis in the wild.Source: §7 · p. 13