An Improved Selective Facial Extraction Model for Age Estimation

In this paper, we propose an improved end-to-end learning algorithm to address the aggregation of multiclass classification and regression for age estimation by using deep Convolutional Neural Networks (CNNs). Inspired by Soft Stagewise Regression Network (SSR-Net), we take residual units embedded with channel and spatial feature response correlation values and dynamically adopted the kernel corresponding to these feature maps into our model. In addition, we used weight normalization at each layer by using input samples at the beginning of training and this weight normalization was beneficial both in terms of accuracy as well as training time. We validate the proposed model based on benchmark datasets and compare the MAE with seven other mainstream networks. The results reveal that our model achieves an improved performance. Our contribution is an updated algorithm for age estimation by adopting the latest attention and normalization mechanisms for balancing the efficiency and accuracy of the proposed model.