Przetwarzanie wstępne i analiza obrazu na użytek lokalizacji twarzy

W artykule zaprezentowany jest algorytm automatycznej detekcji twarzy w obrazie statycznym. Detektor ma osiągać najwyższą skuteczność przy znajdowaniu twarzy możliwie niepochylonych i patrzących na wprost kamery. Wielkość wykrywanych twarzy musi być (z pewnymi odchyleniami) zgodna z rozmiarem twarzy zawartych na obrazach zastosowanych do uczenia klasyfikatora. Obrazy wejściowe mogą być kolorowe lub czarno-białe. Nie ma limitu co do liczby twarzy znajdujących się na obrazie.

The aim of this work is to design and implement a face detection algorithm in static images. The detector have to achieve the best results in finding possible not inclined faces of people looking directly at the camera. The authors have proposed an algorithm which operation is based on the appearance (features) of the face. Block diagram of the proposed face detector is given in Fig. 1. In the first stage, the image containing the face is subjected to preprocessing in which normalization is the most important. Normalization aims to unify a variety of analyzed images. We have used here a conversion of colors to gray levels and stretching and equalization of image histogram. Thus prepared image is processed by the appropriate face detection algorithm, which consists of pre-selection and classification. In order to train the classifier the authors created a database of images consisting of two major categories: containing faces and do not contain faces. As a collection of images that include faces there have been used Olivetti DB ORL database [1]. Final processing step is to get rid of the multiple detection of the same faces. As a result of the algorithm we obtain the location of all faces in the input image (Fig. 4). The size of detected faces should be (with some variations) in accordance with the size of images used to train the classifier. Input images can be color or black and white. There is no limit to the number of faces in an image.