Weighted Heterogeneous Learning for Deep Convolutional Neural Network Based Facial Image Analysis

Recognition of facial attributes such as facial point, gender, and age has been used in marketing strategies and social networking services. Marketing strategies recommend the goods, that are supposed to matches the needs of potential clients. Various social networking services based on facial recognition techniques have recently been developed that can estimate age from a facial image with a high accuracy. To recognize multiple face attributes, we need to train the independent classifier for each type of facial attribute recognition, such as facial point detection, gender recognition, and age estimation. Active appearance model (AAM)[1] and conditional regression forest (CRF)[2] are common approaches for facial point detection. Additionally, age estimation and gender recognition are classified by a support vector machine (SVM) or decision tree using facial point or a local binary pattern (LBP) features [3][4]. With the increasing of deep learning, the deep convolutional neural network (CNN) [5] has become a common classifier for facial point detection [6][7][8], age estimation [9][10], and gender recognition [11][12]. Conventional heterogeneous learning has used the mean squared error function for regression tasks and the cross entropy error function for recognition tasks during the training process. The error ranges of mean squared error function and cross entropy error function are noticeably different. Therefore, we integrate the error range from 0 to 1 by exchanging cross entropy error function for mean squared error function for recognition tasks. However, if we integrate the training error functions, difference of training error is occured, as shown in Fig. 1(a). This difference of training error is occured by difference between label value of regression task and recognition task. The label value of regression