CNN-based Age Estimation from Face Images. Master's Thesis

Age estimation from face images has been a popular and challenging task in computer vision in recent years and is already applied in real-world scenarios such as surveillance, security control, and biometrics. This thesis focuses on deep learning approaches, in particular Convolutional Neural Networks (CNNs), by evaluating and comparing different CNN-based age estimation models trained and tested on the UTKFace dataset. First, relevant literature on the task of age estimation, deep learning techniques, and Explainable AI (XAI) methods used throughout this study are explained. Next, the implementation of CNN models with Depthwise Separable Convolution (DSC) and Pixel Difference Convolution (PDC) layers, treating age estimation as both a classification and a regression task, is explained in detail. In addition, the usage of different methods that deal with data imbalance and the usage of XAI methods is described. The experimental results suggest that treating age estimation as a regression task consistently outperforms classification. Although DSC and PDC layers did not reduce the mean absolute error (MAE), they significantly lowered the number of trainable parameters, resulting in higher computational efficiency, but at the expense of lower accuracy. By using Focal Loss to deal with the imbalanced data, the performance of the model showed further improvements of the MAE. Training on an up-sampled image dataset with the Synthetic Minority Oversampling Technique (SMOTE) led to small improvements of the base model. Other methods used, such as SMOTE on extracted features and Online Hard Example Mining (OHEM), did not yield any improvements. Comparison of XAI methods, such as Saliency Maps, Integrated Gradients, Grad-CAM, and Occlusion, showed that the age estimation model focuses primarily on age-relevant facial features such as wrinkles and facial contours. Among the tested methods, Integrated Gradients provided the most detailed explanations of the models decision-making process.