An enhancement of age and gender classification accuracy with hybrid handcrafted and deep features using hierarchical extreme learning machine / Mohammad Javidan Darugar

• Main Author: Mohammad Javidan , Darugar
• Format: Thesis
• Published: 2020
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://studentsrepo.um.edu.my/14493/2/Mohammad_Javidan.pdf; http://studentsrepo.um.edu.my/14493/1/Mohammad_Javidan.pdf; http://studentsrepo.um.edu.my/14493/

Age and gender classification are some of the essential algorithms that have many use cases in our everyday life. For example, in robotics, field robots can interact with a human base on their gender in data analysis, to have statistics about age and gender of audiences in social events, YouTube video analysis, and many other applications. In this research, we have addressed limitations in deep neural networks, which by overcoming this limitation, we can gain better accuracy and performance. Our study has several other possible applications which are not limited only to age and gender classification. This dissertation is about a high-performance method for age and gender classification in captured facial photos, which is employing deep network architectures as the primary basis of our architecture. We have employed branches of deep learning, such as convolutional neural networks and autoencoders. We have also used Hierarchical Extreme Learning Machines to avoid significant time consumption and to overcome limitations that most conventional deep networks have. We have addressed the problem of using Softmax as the classifier, which usually leads to less performance and accuracy due to its limitation, which is only able to classify linearly separable data. Our proposed architecture consists of two main categories of feature extraction and learning methods which are namely supervised and unsupervised. We have investigated two supervised feature extraction techniques and a deep feature extraction technique to extract unsupervised features to judge the influence of each one on the accuracy of our proposed model. The supervised methods that are examined in this study are Histogram of Oriented Gradients or HOG and Action Units (or AUs). For unsupervised feature extraction techniques, we have employed a deep neural network which is known as convolutional neural network (CNN). CNNs use shift-invariant filters to make discriminative features inside neural networks. One of the very potent tools is convolutional neural networks, which is used for obtaining useful features that also are useful for unknown classes. In our research, we figured out that some of our features are high, and some are very low in dimension. So to combine these groups of supervised and unsupervised features with different dimensions, we have used multiple autoencoder neural networks to join, reduce, and encode all employed feature maps into a single feature vector. Hierarchical-ELM, which is a branch of Extreme Learning Machines, is adopted to classify the final feature vector. Toward this research, we have analyzed the result of our proposed work with state of the art, and also related works are explained to illustrate our significant improvement for age and gender classification in facial images. Our gains are in both accuracy and performance. Regarding performance, we have achieved faster training and testing process. Since we are dealing with large datasets of facial images, therefore, speeding up these steps can influence a more reliable solution.