Investigation of Facial Age Estimation using Deep Learning

Age estimation from facial images has drawn increasing attention in the past fewyears. This thesis project performs the age group classification of facial imagesacquired in in-the-wild conditions using deep convolutional neural networkstechniques. The aim of this thesis project is to propose a novel convolutional neuralnetwork architecture with the intention of investigating the viability of using it inseveral age estimation tasks. First, the proposed convolutional neural network istrained on the preprocessed data sets, and it is found that the network has anobvious overfitting problem. The experiments regarding data augmentation and L2regularization techniques are performed to handle overfitting problems. It is foundthat L2 regularization has a better performance in this task. Second, the viability ofusing the proposed model on differently grouped ages is investigated. It shows that asthe number of age classes increases, the accuracy of the proposed model decreases.Furthermore, downsampling and weight adjustment are investigated to handleimbalanced data. It is found that weight adjustment is more effective thandownsampling to handle imbalanced data in the task. Last but not least, the proposedmodel is compared to some existing state-of-art models on the same age estimationtask, and it is found that the model outperforms LeNet, AlexNet, but fails to beatGoogLeNet in weighted average accuracy, macro average accuracy and one-offaccuracy evaluation metrics