EE368 ProjectFACE DETECTION AND GENDER RECOGNITIONMichael Bax, Chunlei Liu, and Ping Li26 May 20031 IntroductionThe full face detection and gender recognition system de-scribed here is made up of a series of components con-nected in both serial and parallel, as illustrated in Figure 2.The successive stages are explained in detail in the bodyof this report.2 Colour-based segmentationThe first step in this face detection algorithm is that ofcolour segmentation. The goal is to remove the maxi-mum number of non-face pixels from the images in or-der to narrow the focus to the remaining predominantlyskin-coloured regions.2.1 Colour space selectionThe first step in performing colour-based segmentationis choosing an appropriate colour space in which to op-erate from the wide variety of choices such as RGB,HSV, CMYK, YCbCr, etc [1]. Of these, RGB (red-green-blue) and HSV (hue-saturation-value) have been the mostwidely used. Figure 1 illustrates the geometries of the twospaces.By way of example, HSV representation has certain ad-vantages over RGB when it comes to face detection. AsGarcia et al [2] note, skin colours are sensitive to the light-ing condition. In the RGB space, each of the three com-ponents may exhibit substantial variation under differentlighting environments. In HSV space, however, the hueand saturation components are virtually unchanged.Figure 3 shows the histograms of RGB component val-ues of both face and non-face pixels overall seven traininginput images. Similarly, Figure 4 shows the histograms ofthe same images in HSV space, where the S componentand in particular the H component are well-clustered forface-pixels, while H and S are spread over a wide rangefor the remainder of the image. This observation favoursusing an HSV colour space if only a simple thresholdingcolour segmentation is desired.GRBMagentaBlackWhiteCyanYellowGreenRedBlue(a) RGB modelSHVGreen YellowMagentaBlueBlackWhiteRedCyan(b) HSV modelFigure 1: The RGB and HSV colour models.1MaximaselectionAverage facetemplate matchingMulti−inputanalysisConservative faceregion isolationRegionclassifierRegionclassifierAggressive faceregion isolationRegionlabellingMorphologicaloperationsPrecision face centretemplate matchingMaximaselectionMaximaselectionIndividual female facetemplate matchingGreyscale pyramidcompositionMorphologicaloperationscolour−segmentationHistogram−basedFace location and genderPhotographFigure 2: The face detection system schematic.2The same is true of the YCbCr colour-space, but theclustering does not quite as readily lend itself to projectiononto the axial dimensions, and hence requires a somewhatmore complicated 2D basis for segmentation.Manual segmentations of this type suffer as a conse-quence of approximating a complex bounding region us-ing simple geometric relations. Given that the unknownimages were taken in similar lighting and with similarequipment to the training images, a probability-based seg-mentation may be used.In this case, the choice of spatial domain is not as sig-nificant. However, the fundamental characteristics of acolour space can complicate processing — for example,the non-Cartesian nature of the HSV colour cone indicatesthat non-uniform quantization may be appropriate.2.2 Probability-based classificationDue to the large number of pixels in the training images,there is enough data to create a reasonable estimate of theunderlying probability density functions for both face andnon-face skin colours. LetfX(x|X ∈ Φ)be the colour space probability density function for a pixelvector X in the set of face pixels Φ; the vector componentsXirepresent the colour components — R, G and B in thiscase. Similarly, letfX(x|X /∈ Φ)be the probability density function for non-face skin pix-els. These two density functions can be estimated fromthe empirical distribution of the pixels in the training im-ages.Conversely, let the probability that a given colour pixelis part of a face bepX(X ∈ Φ|X = x).The Bayesian formula [5] givesR =pX(X ∈ Φ|X = x)pX(X /∈ Φ|X = x)=fX(x|X ∈ Φ)fX(x|X /∈ Φ)×πΦπ¯Φ(1)where πΦand π¯Φare the prior probabilities of a randomly-selected pixel falling in a face or the background, respec-tively. Without further information, the ratio ofπΦπ¯Φcan beestimated from the ratio of the total number of face pixelsto the total number of non-face pixels.For convenience, Equation 1 may be reformulated aslogR = logpX(X ∈ Φ|X = x)pX(X /∈ Φ|X = x)= logfX(x|X ∈ Φ)fX(x|X /∈ Φ)+ logπΦπ¯Φ. (2)If the prior information is known, the classification rulewould normally be:X ∈ Φ R ≥ R0= 1, i.e. logR ≥ logR0= 0X /∈ Φ otherwise.However, the prior information is not necessarily avail-able; conversely, in many situations it is desirable to ad-just the classification threshold R0. For example, since inthis approach colour-based segmentation is used for sub-sequent face detection, it is desirable to bias towards mis-classifying a non-face pixel as a face pixel rather than thereverse. Consideration of these factors yields the follow-ing classification ruleX ∈ Φ logfX(x|X ∈ Φ)fX(x|X /∈ Φ)+ α ≥ 0X /∈ Φ otherwise, (3)whereα = logπΦπ¯Φ− logR0(4)is the undecided parameter to be chosen.A range of values for α for image pixel classification inboth RGB and HSV space were evaluated, the results ofwhich are shown in Figure 5. The total classification erroris composed of false positive error (mis-classifying non-face pixels as face pixels) and false negative error (mis-classifying face pixels as non-face pixels).The two sub-figures are quite similar. The false nega-tive error increases with α, while false positive error de-creases with increasing α. The total error reaches mini-mum at α = 3. This optimal value of α is supported bythe observation that, over all 7 training images, the ratioof the total number of face pixels to non-face pixels isroughly116; log116= −2.77.Because this colour-based segmentation is only the firststep in the larger face detection algorithm, it is best to re-tain as many face pixels as possible, minimizing false neg-ative error (even at the cost of including additional non-face pixels), leading to the chosen value of α ≈ 2.Although the performance differences between RGBand HSV colour space are quite small, RGB is nonethe-less superior [3], and that was the colour space chosen forthis algorithm.In practice an empirical histogram derived from man-ual image segmentation is somewhat