Estimating of age and gender information from face images with multiple machine learning methods

The applications that facilitating the life with computer systems are quite important. Pattern recognition which is being used commonly in security systems has become an increasingly important field of study. This field of study is being used for many purposes such as character estimation from handwriting, medical diagnosis from disordered voice and physical appearance, controling of unmanned aerial and ground vehicles. There are also many fields of studies like human, character, age and gender estimation from face images. Age and genders are estimated by using face images in this thesis. Essentially age and gender estimation are two different problems. In this study both of these problems are considered as a single. Determining the minimum feature count to describes these problems and getting maximum accuracy values in minimum calculation times after classification are defined as the main target. In this study the Viola-Jones algorithm, which is used commonly in literature, was used for detection of face patterns. This algorithm is successful only when the face area has a vertical angle of 15 degrees or less. In this study, an improved version of the Viola-Jones algorithm was developed which is able to detect face patterns with any vertical angle depending on a rotation parameter given by the user. The rotation angle was presumed as 15 degrees in this study. Thus, face patterns which have a vertical angle of 30 degrees was searched in images. Two different feature extraction method was used as hybrid for feature extraction from face patterns. Acquired dataset dimensions have been reduced by using feature selection algorithm. According to the classification results that have been acquired by using the proposed method on 800 images from color FERET database; the age estimation was completed in 2,63 seconds with %93,1250 accuracy, the gender estimation was completed in 2,86 seconds with %96,8750 accuracy. All of these results obtained by using 20 fold cross validation method. High accuracy values and short calculation times show that the proposed method is usable in embedded systems. It is observed that the accuracy results obtained by the proposed method are higher than many results in literature.