Gender Classification of Human Faces Using Class Based PCA

Gender classification is a binary classification system where system has to assign a given test image to one of the two classes (male or female). The gender classification system with large set of training data normally gives good accuracy. But to achieve good accuracy with small training data is a difficult task. This paper proposes an algorithm for gender classification with small training data and it gives good accuracy even with one image per person for training. The system contains mainly two parts: feature vector generation and classification. Feature vector generation is done with PCA (Principal Component Analysis ). Generally all training images are organized as a columns of a matrix and then PCA is applied to generate feature vectors of those training images. To reduce the computational complexity, PCA is separately applied to each individual training class. The paper proposes a new approach of classification where, after a given test image is reconstructed with different Eigen coordinate systems, lowest MSE(mean square error) between given image and reconstructed image, indicate the output class for that image. The proposed method is also compared with nearest neighbor classification using different similarity criteria such as Euclidean, Manhattan, Chebyshev, Canberra, Cosine Correlation and Bray-Curtis distance. These all algorithms are applied on two databases. Indian face database (664 images) and local database (1000 images). Results show that the proposed method significantly improves the overall classification accuracy.