Abstract:
Using the UTKFace dataset, this study presents a novel design of Convolutional
Neural Networks (CNNs) to accurately identify age and gender solely from facial
imagery. A well-structured CNN model from scratch utilizes RGB images of the
dataset by employing fully connected and convolutional layers, along with Adam
optimizer for optimal performance and outstanding accuracy in gender prediction.
Similarly, a comprehensive CNN model from scratch is also employed for age
prediction after hyperparameter tuning, correctly identifying the distinct age
categories with the implementation of cross-entropy loss and L2 regularization.
Another model of Machine Learning called Support Vector Machine (SVM) has also
been developed that utilizes Histogram Oriented Gradient (HOG) and Gabor Filters.
The study focuses on dataset preparation, data preprocessing, model training,
hyperparameter refining, and optimization and evaluation metrics. It also ensures the
flexibility of the models through training, validation and testing on the popular UTK
Face Dataset. With precise dataset curation and feature extraction, the techniques of
ML and DL provide a solid base for the integration of the trained model to an
application software interface to predict age and gender from facial images of the
individuals in real time. Additionally, it also offers further opportunities for future
developments and applications in other industries that depend on facial analysis
information.
