Abstract:
Face plays a vital role in the identification or recognition of a person. It has significant applications for the enforcement of law and order, for training purposes in the education department, for creating a specific character in the entertainment industry, etc. But sometimes the facial image of a particular person is not available, and a sketch of that person is required. Because, if we are able to acquire a sketch of that person, then it can be used to generate realistic face images. Traditional face generation systems required professional sketches to generate realistic images. Professional artists are needed to draw those sketches but there is a possibility that professionals are not always available for the accomplishment of this task. There is a need for an automated system where even novice users can also draw a sketch of that person on the basis of certain information. This research study leads toward a solution where a novice user can also draw a sketch of a particular person. A database of basic or rough sketches is developed to accomplish this task. These sketches are used to generate realistic face images. Generative Adversarial Network (GANs) models are used for this sketch-to-photo translation. We have used two variants of conditional Generative Adversarial Networks (cGANs) that perform better for our problem domain. Those models are the pix2pix-GAN model and Cycle-GAN that we used during this thesis work. Both qualitative and quantitative analyses of generated images have been conducted to evaluate the performance of our models. Finally, we have used different visualization techniques to better understand latent space. Results of both analyses and visualization of latent space show that our proposed methodology is performing very well for this problem domain
