$1(85$/1(7:25.0(7+2’)25$&&85$7()$&(’(7(&7,21 21$5%,75$5<,0$*(6 D. Anifantis, E. Dermatas and G. Kokkinakis Wire Communications Laboratory, Department of Electrical & Computer Engineering, University of Patras Patras, Kato Kastritsi, 26500 HELLAS Email: anifanti, dermatas@george.wcl2.ee.upatras.gr ABSTRACT In this paper we present a neural detector of frontal faces in gray scale images under arbitrary face size, orientation, facial expression, skin color, lighting conditions and background environment. In a two-level process, a window normalization module reduces the variability of the features and a neural classifier generates multiple face position hypotheses. Extended experiments carried out in a test-bed of 6406 face images, have shown that the face detection accuracy is increased significantly when non-linear and probabilistic illumination equalizers pre-process the sub-images. Moreover, better results can be achieved in case of training the neural detector using positional and orientation normalized face examples. In this case the neural face detector has the capability to locate both position and orientation of a face. In the multiple face position hypotheses generated by the proposed neural method, 98.3% detection accuracy, the highest reported in the literature, was measured. 1. INTRODUCTION Human face detection methods have been studied for more than 20 years. The face identification problem in poor illumination conditions, different perspective variations, and background environment becomes a great challenge in real-life applications [1-3,9]. Faces are similarly structured with the same features arranged in roughly the same spatial configuration. Nevertheless, even among images of the same person's face, significant geometrical and textural differences are met due to changes in expression, illumination conditions and the presence of facial makeup. Therefore, traditional template matching and geometrical object recognition methods tend to fail by detecting faces in arbitrary lighting, a change in light source distribution can cast or remove significant shadows from a particular face, and in different back- ground conditions. Neural detectors have been pro- posed recently and the experiments give extremely satisfactory results [4]. In particular, advanced training methods have been used to incorporate the wide distribution of the frontal view of face patterns in neural network knowledge [4,5]. An efficient solution to the selection problem of non-face images has already been presented aiming at minimizing the false acceptance error [4]. This technique reduces the number of falsely recognized faces in arbitrary images, a critical factor for developing practical face recognizers. The most accurate face detector has been reported recently giving recognition rates up to 96% by using a neural network method [4]. In this paper we present and evaluate an accurate neural detector of frontal view faces on arbitrary images. In a two-level process, the window normalization and the NN classification, a number of image window hypotheses is generated and charac- terized as face regions (FRs). In the window normali- zation process, each sub-image is transformed into a 30x30-pixel image and illumination filters minimize the dispersion of the illumination phenomenon. The neural network classifier detects the presence of a face by mapping the input vector space into the two- dimensional space of the neural network output. Each output value represents the belief level of the neural network knowledge on the presence or absence of a face. The proposed structure of the Multilayer Perceptron neural network increases the computational complexity but it faces the problem of detecting the position and the orientation of faces with different features (skin color, presence of a moustache, bear, glasses, etc). The FRs are detected by scanning each image using window analysis in an overlapping mode. The proposed method has been evaluated in a test- bed of 6406 faces (the Olivetti, and Yale databases, a subset of the FERET database, a collection of face im- ages downloaded from the WWW and a set of 350 texture and 220 scenery images. The experiments have shown that the proposed FR detection method’s accuracy is improved significantly in case of incorporating the non-linear equalizers of illumination and performing orientation calibration of the training faces, giving a face detection accuracy of 98.3%. The rest of this paper is organized as follows. In the next section the proposed method is described in detail. Section 3 presents the face image database used for the evaluation of the method. In the last section the experimental results and the current directions of our research are given.