Abstract:  In order to further enhance the accuracy of age estimation, we proposed a locally adjusted age estimation algorithm based on deep learning and directed acyclic graph-support vector machine (SVM). In the training phase, the SE-ResNet-50 network, pre-trained on the VGGFace2 data set, was first fine-tuned. When it converged, the fully connected layer was extracted, and the vector formed by its end-to-end connection was employed as a representation and further trained multiple one-versus-one SVM. In the testing phase, we first sent the face image into SE-ResNet-50 to obtain a rough age result, then set the specific neighborhood, finally integrated the trained SVM into a directed acyclic graph SVM, and conducted accurate age estimation centering on the global estimation value. In order to show the universality of the algorithm, the results of experiments undertaken in MORPH and AFAD datasets of different races can verify the effectiveness of the algorithm. 

Key words: age estimation, deep learning, directed acyclic graph support vector machine, local adjustment