Flow-chart of the proposed face detection algorithm. 

Context in source publication

Context 1

... detection from images is a key problem in human computer interaction studies and in pattern recognition researches. Detecting faces is the first crucial step for face recognition and as a consequence it has significant applications such video surveillance, human computer interface, and face image database management. Many different approaches for face localization are published in the literature. All the face detection attempts rely on only one type of algorithm for detection, sometimes followed by other types of techniques for further validation. Most methods use only the luminance component, extracting features as texture, depth, shape, and eigenfaces. There are applied various techniques as learning algorithms, bootstraps, SVM, neural networks and fuzzy methods. A second group of methods added the chrominance information as a validation of the luminance channel technique. The general idea behind them was to confirm through color analysis that the candidate object has face- like color. A third type of algorithms starts from the chrominance information to locate candidate faces, which are then validated by searching other facial features. The necessity of finding suitable facial features for validation is due to the fact that color analysis yields information related to the presence of skin rather than the presence of face. In this paper, we propose a model that belongs to a new algorithm category consisting of the two competing luminance/ chrominance channels. Our model uses two corresponding techniques of finding face candidates and proposes a dual cross-validation of the two branch decisions in an attempt to cover different situations. This cross-validation is different from the simple logical “AND” of the two channel decisions. The color channel uses skin detection based on color histogram, followed by analysis of shape information and ellipse fitting. When the luminance channel has the function of main face detector, it applies the Viola-Jones algorithm [11]. For validation of the chrominance detection, the luminance channel applies an improved SVM technique for fast detection [4], [8]. When building a system that uses skin color as a feature for face detection, first main problem is to choose the color space. The main purpose of this paper is dedicated to the improvement of face detection rate in color images by applying the color space conversion model previously proposed by Neagoe in [6], [7] for color pattern recognition; this implies the color space conversion from the conventional RGB space into the 3D uncorrelated color space (UCS), using the Karhunen-Loève transform (KLT), equivalent to Principal Component Analysis (PCA) in the color space. The paper is structured as follows. Second section presents the flowchart of the proposed algorithm. Third section is dedicated to the experimental results, while the fourth section contain concluding remarks. The present model has three main sources. Firstly, for chrominance channel, this model is inspired by the approach of Sobottka and Pitas [10] for face detection, based on the observation that human faces are characterized by their oval shape and skin-color, also in the case of varying light conditions. It applies color segmentation in HSV color space and it is followed by analysis of the shape information by ellipse fitting. Secondly, for luminance channels, this paper applies either Viola-Jones face detection algorithm [11] or the fast SVM of Kienzle et al [4]. Thirdly, to the aim of improving detection performance, we have applied the color space model proposed by Neagoe [6], [7] consisting in conversion of the conventional RGB space into the 3D uncorrelated color space (UCS), using the Karhunen-Loève transform (KLT). On the other side, face detection algorithms presented in literature belong to one of the following three categories: (a) methods using luminance information; (b) algorithms adding the chrominance decision as a validation of the luminance decision; (c) methods based only on the chrominance information for face candidate detection. We further propose a face detection model (see Fig. 1), thus initiating a new category. It has two competing chrominance/luminance branches and two specific face candidate finding techniques that cross-validate their decisions each other; the final decision is a result different of a simple logical intersection of branch decisions. The main branch of the flow-chart corresponds to the locating of the candidates by chrominance analysis and to the cross- validation by a technique relying on the luminance channel. We have covered several color spaces, to comparatively evaluate the algorithm detection performance. We further consider the Karhunen-Loève transformation (KLT) for conversion of the RGB space into a 3D Uncorrelated Color Space (UCS) . For comparison, one evaluates the application of the YCbCr and HSV color spaces. Decorrelation of the Color Space. In order to improve the color pattern recognition performances, we shall further apply the color space conversion model proposed by Neagoe in [6], [7]. Consider the color pixels in a given image as 3D vectors P(x, y) = [R(x, y) G(x, y) B(x, y)] t (1) where R(x, y), G(x, y) and B(x, y) are the red, green and blue components of the pixel of coordinates (x, y). We assume that color images exhibit features that can be useful in the conversion from a 3D full color space representation to the 3D uncorrelated color space (UCS); as we further prove, this transformation improves correct pattern recognition score. For color conversion, we have chosen the Karhunen-Loève transformation (KLT), also known as Principal Component Analysis (PCA), to eliminate the correlation of the R, G, and B color channels. The interesting fact in our case is that one preserves the space dimensionality (from 3D RGB to 3D UCS). The reason of applying PCA here is not to reduce the space dimension but only to eliminate the component correlation. To deduce the KLT matrix, one firstly computes the covariance matrix of the color pixels (represented as 3D vectors). Then, one computes the eigenvalues of the covariance matrix. Finally, we deduce the three eigenvectors. Thus, one obtains the KLT matrix ...