Fusion of Hand-crafted Descriptors with CNN-based Features for Facial Age Estimation

Age estimation from facial images is an important application of biometrics. In contrast to other facial variations like occlusions, illumination, misalignment and facial expressions, ageing variation is affected by human genes, environment, lifestyle and health which make age estimation a challenging task. In this thesis, we propose three new age estimation systems for automatic facial age estimation. These systems utilize different type of feature descriptors, varying from hand-crafted ones to automatically learned features and combine them in different level of information fusion. In the first proposed system, an integration of different feature extraction algorithms is utilized. This integration is performed by using two-level fusion of features and scores with the help of feature-level and score-level fusion techniques. In our proposed method, the advantage of using different types of features such as biologically-inspired features, texture-based features and appearance-based features is used. Feature-level fusion of biologically-inspired and texture-based methods is integrated into the proposed method and their combination is fused with an appearance-based method using score-level fusion. The second proposed system exploits multi-stage features from a trained Convolutional Neural Network (CNN), and precisely combines these features with a selection of age-related hand-crafted features. This method utilizes a decision-level fusion of estimated ages by two different approaches; the first one uses feature-level fusion of different hand-crafted local feature descriptors for wrinkle, skin and facial iv component while the second one uses score-level fusion of different feature layers of a CNN for age estimation. In the third system, we propose a new architecture of deep neural networks namely Directed Acyclic Graph Convolutional Neural Networks (DAG-CNNs) for age estimation which automatically combine multi-stage features from different layers of a CNN. This system is constructed by adding multi-scale output connections to an underlying backbone from two well-known deep learning architectures, namely VGG-16 and GoogLeNet. DAG-CNNs not only fuse the feature extraction and classification stages of the age estimation into a single automated learning procedure, but also utilize multi-scale features and perform score-level fusion of multiple classifiers automatically. Experiments on the publicly available Morph-II and FG-NET databases prove the effectiveness of our novel method.