earticle

영어

So far, several methods have been proposed for gender estimation based on training processes, using several images per person. In this paper, we introduce a new gender estimation method based on single image per person. First, reference images, called general faces, are constructed by averaging male and female images in the dataset respectively. Then, important face features such as eyes and mouth are highlighted by Wavelet fusion of Sobel Detector, and Laplacian of Gaussian images. The highlighted image is segmented into seven partitions. In the obtained image, those segments like eyes and eyebrows with more variations (containing more information) play more significant role in the gender estimation process. Weight of each segment is proportional to its entropy. Finally, DCT coefficients of each segment in the input facial image are compared to their counterparts in general faces. Considering weight of each segment and fuzzy rule sets, the input image is assigned to the closest class. The proposed method is evaluated on FERET and AR databases. Experimental results show that the proposed method is tolerant of variations such as facial expression, occlusion, and rotation.