Age and gender classification from ear images

Ear is one of the important biometric traits. Because, the human ear can be easily obtained and is unique for each individual. Nowadays, ear biometrics has become a very popular research topic in biometric research area. In this thesis, we present a detailed analysis on age and gender classification from ear images. Although there have been several studies on gender classification, to the best of our knowledge this thesis is the first work about age classification from ear images. In this work, we have presented both geometric based representation and appearance based representation for age estimation and gender classification from ear images. In geometric based representation, firstly, we have identified several landmark points both on the ear and on the profile face. Then, we have calculated distances between them and have also calculated rectangle area and polygon area of the ear to use as features with machine learning algorithms. We have employed support vector machines (SVM), linear support vector machines, neural network, random forest, and logistic regression with geometric features and we have analyzed the age and gender prediction results. We have also performed appearance based representation. For this, we have utilized VGG-16 and ResNet-50 deep convolutional neural network architectures with ear images. Moreover, we have presented two-stage fine-tuning strategy and data augmentation approaches to improve classification peformances. For domain adaptation method, first of all, pretrained CNN model on ImageNet dataset has been fine-tuned on the large-scale ear dataset which is called as Multi-PIE ear dataset. With this way, we have aimed to adapt the CNN model to the ear domain. Later, we have fine-tuned this model on our target datasets which are Istanbul University dataset, FERET, UND-F and UND-J2 respectively. We have also performed data augmentation technique on all datasets for producing more data and increasing the variance of the datasets in order to improve the classification performance of the deep CNN models. Besides, we have benefited from center loss, large-margin softmax loss (L-Softmax), and angular softmax loss (A-Softmax) in deep CNN model. With these loss functions, we have obtained more distinguishable features and more accurate results. While Istanbul University dataset and FERET dataset have been employed for both age and gender classification, UND-F and UND-J2 datasets have only been utilized for gender classification experiments due to lack of age information of the subjects. On Istanbul University dataset with geometric features, we have obtained 38.5% classification score by linear SVM for age classification, and we have achieved 63% classification performance by linear SVM for gender classification task. On the other hand, for appearance based representation, we have reached 54.41% age classification accuracy by VGG-16 CNN model and 100% gender classification accuracy by both VGG-16 and ResNet-50 using large-margin softmax loss. For age classification experiments on the FERET dataset, in the 3 class setup, we have obtained 85.71% result with ResNet-50 and center loss. In the 5 class setup, ResNet-50 and VGG-16 deep CNN architectures have achieved 69% classification accuracy with large-margin softmax loss. Moreover, for gender classification, we have obtained 98.33%, 98.33%, and 99.16% classification performance on FERET, UND-F, and UND-J2 datasets respectively. All these appearance-based results have been obtained with data augmentation and domain adaptation techniques. Finally, all results indicate that ear can provide useful information for age and gender classification. However, age estimation requires further work.