Age Estimation via Fusion of Depthwise Separable Convolutional Neural Networks

In this paper, a deep Convolutional Neural Network CNN based system, called Depthwise Separable Convolutional Neural Network (DSCNN) fusion system, for human facial age estimation is presented. This system includes following four stages. In the first stage, a data augmentation procedure is utilized to enrich the dataset. In the second stage, a pre-trained deep CNN model is fine-tuned for the gender classification task. For the third stage, three newly designed DSCNN age estimators are utilized to conduct gender-specific age estimation for gender grouped facial images from previous stage. The architectures of these three deep DSCNNs are constructed to lower computation complexity. In the last stage, estimated ages from three DSCNN age estimators are fed to the fuser to boost the overall age estimation performance. In the experimental results, on four benchmark datasets, IMDB-WIKI, MORPH-II, and ChaLearn LAP Apparent age V1 and V2, the proposed system demonstrates a significant performance improvement over the state-of-the-art deep CNN models and methods.