Child-Face-Generation

Child face generation is a computer vision problem in which the goal is to synthesize realistic images of a child given images of its parents. We present a model for this problem based on Deep Convolutional Generative Adversarial Networks (DCGANs). Key challenges in this domain include limited datasets with high dimensional input spaces and the multi-modal nature of the target distribution. We demonstrate convincingly that GANs have a unique ability to capture the latter feature, while use of state-of-the-art training techniques and architecture optimizations allow us to mitigate the impact of the former. As a baseline, we use a simple supervised model that minimizes RMSE with respect to target images. Qualitatively, the clarity and diversity of images reflect advantages of the GAN model when compared to the base model. Quantitatively, we find that after training the discriminator correctly classifies GAN-generated images as fake with a rate of 71.1%, compared to 99.5% classification accuracy on the images generated by the baseline model, suggesting an objective basis for our observations.

Generator Model Architecture:

Discriminator Model Architecture:

Results Table: