Research Of Gender Recognition Based On Static Facial Images

With the rapid development of globalization and network, the identity recognition technique in everyday life is urgently demand. In the human biological characteristics, face features being easy get to make it widely used in identity recognition. Gender is one of the important face characteristics, and its recognition has become one of the hotspots.Bo W(bag of words) model is widely used in the field of document classification methods. In recent years, some researchers have tried to apply it to image classification, such as thebag-of-words visual vocabulary and visual dictionary. Since Bo W model is simple and effective features, we deeply study the Bo W model in this paper.Gender recognition is a typical binary classification problem. First, this paper introduces various factors that affects the gender recognition. Then the common face database and the pre-processing methods for facial images are given.Based on the deeply study of Bo W model, Adaboost algorithm and Naive Bayesian algorithm, this paper proposes a modified method of Bo W model based on Dense SIFT and NBC using spatially-constrained similarity measure. Traditional sparse SIFT features need to build the Gaussian scale-space and difference of Gaussian scale space during the process of extraction, so constructing the scale-space require more time and space, while the extraction process of Dense SIFT feature is not required to build the space. Dense SIFT gets samples images uniformly. In the actual process of feature extraction, it also has strong adaptability to different backgrounds and improves the efficiency of feature extraction.Finally by using the images on the CAS-PEAL face database, under different face changing conditions, such as posture, facial expressions, lighting and accessories, etc. the experiment is carried out with Na?ve Bayes classifier and Adaboost classifier combined with Bo W feature extraction model based on Dense SIFT and Sparse SIFT. The experiment results show that the model based on Dense SIFT's Bo W combined with NBC improve the accuracy of 2% than several other recognition method.