A Study Of Gender Classification And Age Estimation Based On Face Image

Recently, Biometrics has been applied into pattern recognition including face recognition, fingerprint recognition, iris recognition, palm print recognition etc. Compared with other biometric personal authentication technologies, face recognition technology is natural, convenient and non-contact. These advantages make it be applied widely in security surveillance, verification, man-machine conversation and so on. Gender classification and age estimation is an attempt to give the computers the ability to discriminate the gender information and estimation age from a face image. Gender classification and age estimation has become the hotspot of computer vision and pattern recognition for its important application prospects in identity recognition, man-machine interface, video index and robot vision and so on. However, it is still one of the most challenging problems in the fields of computer vision.Since feature extraction algorithms used by gender classification and age estimation just need to extract the features of face area, face detection is a necessary step in automatic gender classification and age estimation systems. Moreover, the accuracy of face detection would affect the effectivity of feature extraction. So multi-view face detection method based on Adaboost learning algorithm is applied, and its cascaded classifier structure and pyramid architecture both improve the speed of detection. The experimental results show the face detection method is rapid and effective.In order to normalize the size of face image, an eye location method based on Adaboost and fast radial symmetry transform is proposed. Firstly, fast radial symmetry transform can search feature points rapidly. Secondly, an eyebrow region detection algorithm based on Adaboost is presented to reduce the range of searching eyes, which can decrease the influence of other feature points on eye location. Finally, the precise location of eyes can be obtained by using pupil model and geometrical features of eyes. Furthermore, the results of eye location are used to initialize active appearance model, which is applied to locate facial feature points so as to extract the local features of face images.Gender classification and face recognition both need to find effective and stable facial features, but ultimately identify the different patterns. In this paper, we compared those methods which used in gender classification, including local binary pattern, neural network, Adaboost, support vector machines (SVM), and image pixels as input. We present a systematic study on gender classification with automatically detected and aligned faces. We experimented with different combinations of automatic face detection, face alignment and gender classification, and show the classification rates for different face images sizes.Based on the type of features used, previous studies can be broadly classified into two categories: appearance feature-based (global) and geometric feature-based (local). The former finds the decision boundary directly from training images while the latter is based on geometric features such as eyebrows thickness, nose width, etc. The local and global features supplement each other under some conditions. In this paper, a novel gender classification method based on frontal face images is presented. In this work, the global features are extracted using AdaBoost algorithm. Active Appearance Model locates 83 landmarks, from which the local features are characterized. After the fusion of the local and global features, the mixed features are used to train support vector machine classifiers. This method is evaluated by the recognition rates over a mixed face database containing over 21,300 images from 4 sources (AR, FERET, CAS-PEAL, WWW and a database collected by the lab). Experimental results show that the hybrid method outperforms the unmixed appearance- or geometry-feature based methods and achieve a classification rate over 90%.In the past, most computational models of gender classification use global information (the whole face image) giving equal weight to all areas of the face, irrespective of the importance of internal features. Intuitively, we argue that smaller facial regions, if judiciously selected, would be less sensitive to expression variations and may lead to better overall performance. We evaluate the significance of different facial regions for gender perception. Our work on gender classification is one of the first attempts to report a detailed evaluation of the significance of different facial regions including the whole face (including hairline), the internal face, the upper region of face, the lower region of face, the left eye, the nose, and the mouth. Considering the significance of facial regions, we propose a fusion-based method, combining the classification results of three facial regions, for improving the robustness to facial expressions.We notice that most computational models of the existing age estimation methods consider only the entire face as a global feature, they do not take into account just what other regions of the face as local features. We propose a novel method for age estimation that combines information from multiple facial features for improving accuracy and robustness. The facial features that we consider are the grayscale image of the face, the Gabor wavelet representation of the face, and the eyes. The Gabor wavelet representations are robust for illumination and expressional variability, so are used widely in facial feature modeling in recent years. The eyes are essentially unaffected by beards and mustaches and quite robust to facial expressions and occlusions. Moreover, the area around the eyes was found to be the most significant for the task of age estimation. The idea is to use complementary information for improving overall performance. Further, we propose a fusion method that further improves the accuracy. Each feature provides an opinion on the claim in terms of a confidence value which is calculated by SVMs. The confidence values of all the three features are then fused for final age estimation. The proposed fusion method works quite well and yields a significant improvement in age estimation over that achievable with any single feature.