Fig. 1. Image segmentation. (Left) Original, (Right) image segmented using EDISON.

Fig. 2. Removing specularities. (Left) Original, (Right) after specularities have been removed using the method in this paper.

Fig. 3. An expanding wavefront that takes the shape of a container.

Fig. 4. A specularity can be considered as a 3D surface such as a mountain. Intensity plots of three typical specularities.

Fig. 5. Contour map of specularity. Each contour level can be thought of as a wavefront.

Fig. 6. Plot of contour level versus total region size for some typical specularities. Each contour level has an intensity of 1% less than the previous one. The sharp increase in slope marks the point where the plain (matte) region begins and the mountain (specularity) ends.

Fig. 7. Results of growing wavefronts outward from the peaks of potential specularities. (Left) Original image, (Middle) specular areas detected with the IS mask which are used to determine the initial seed points. (Right) Specular regions obtained by growing outward from the peak of each potential specularity.

Fig. 8. Sometimes the expansion is cut short too early. (Left) Original, (Middle) specular areas detected with the IS mask. A seed point that is barely visible is detected on the forehead. (Right) After growing outwards: the wavefront fails to expand from the seed point to the boundary of the specularity.

Fig. 9. Plot of contour level versus total region size. Very often during the first few iterations the region size increases very minimally. Compare with Fig. 6.

Fig. 10. Clipping the noise line prevents the expansion from being cut short. (Left) Original, (Middle) growing out from the peak of the specularity. (Right) Growing out from the peak of the specularity after first clipping the mesa.

Fig. 11. Detecting specularities. (Top) Original, (Bottom) detected specular regions.

Fig. 12. The IS Diagram. (Top) Original, (Middle) LMSR [1] output, (Bottom) IS Diagram of Luminance Retinex output.

Fig. 13. Binary mask used in [17] (no scale provided in the original).

Fig. 14. Creating the Binary Mask. (Left) Seed points. (Middle) Lines (red) fitted to the extremities of the seed points. The equations of the lines take the form S = mI + b, where m₁ = 1.117, m₂ = 0.875, m₃ = 0.48, m₄ = −0.74, m₅ = −1.24, m₆ = 2.22, m₇ = 0.01, and b₁ = −1.05, b₂ = −0.28, b₃ = 0.48, b₄ = 0.69, b₅ = 1.22, b₆ = 1.69, b₇ = 0.65. (Right) Final Mask. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this paper.)

Fig. 15. Detecting specularities using the IS binary mask. (Top) Original. (Bottom) Specularities thresholded using the IS mask.

Fig. 16. Intensity histograms. (Left) The RUI, (Right) the RUI after gain-offset correction for a typical facial image.

Fig. 17. (Left) Original, (Right) the MBI.

Fig. 18. Coloring inwards. From left to right: as the wavefront boundary is repeatedly colored inwards the specularity becomes smaller and smaller until it is quenched. The detected specular region is red and the contracting wavefront is green. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this paper.)

Fig. 19. That part of the specularity boundary (wavefront) that coincides with a material boundary must be frozen. The rest of the wavefront is colored inwards as shown. Above the specularity is red, the wavefront is green, and the black arrows indicate the directions that the wavefront contracts as it colors inwards. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this paper.)

Fig. 20. Freezing that part of the specularity boundary that coincides with a material boundary. In this example, the specularity shares a boundary with the eyebrow. (Left) Original, (Middle) specularity in red, (Right) result of coloring inwards as per Fig. 19. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this paper.)

Fig. 21. Specularity detection and removal. For each pair of images, the left is the original image, while the right is the result of applying the LMSR [1] and then detecting and removing specularities.

Fig. 22. Illumination compensation applied to a variety of images. From left to right: (A) Original image, (B) illumination compensated image, (C) original image segmented, and (D) illumination compensated image segmented.

Fig. 23. Illumination compensation applied to the two faces in Fig. 22. (A) Woman in 3rd row; (B) RGB histogram of image in (A); (C) woman in 4th row; and (D) RGB histogram of image in (C).

Fig. 24. The skin locus. Skin pixels are purple while the Planckian locus is the black curve. The skin locus is camera specific. (Left) Nogatech camera [30], (Right) Winnov camera [28].

Fig. 25. Generic Skin Locus. The loci of several cameras [26], [27], [28], [29], [30], [31] and [32] were studied to create a skin locus that caters to a generic camera.

Fig. 26. Binary mask of skin pixels. (Left) Original, (Middle) after illumination compensation, (Right) manually obtained binary mask of skin pixels.

Fig. 27. Distribution in chromaticity space of skin pixels within the binary mask. Skin pixels are in red and the boundary of the generic skin locus is black. (Left) Original image, (Right) after illumination compensation. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this paper.)

Fig. 28. Face mask with correctly labeled skin pixels in red and the remaining errors in green. (Left) Uncompensated image, (Right) illumination-compensated image.

Fig. 29. Typical frontal images with varying illumination from the CMU PIE database. Note the slight in-plane rotations, varying eye positions, and non-uniform background.

Fig. 30. Images from Fig. 29 after geometric but not illumination normalization which is illustrated in Fig. 28.

Fig. 31. The experimental process: all images were normalized before recognition was performed.