Image Generation Network Model based on Principal Component Analysis

In the field of Artificial Intelligence, a large and densely annotated dataset is required for training making it a time and resource-expensive task. In this paper, we propose an image generation network model that keeps the training examples at a minimal level. The proposed model gives additional feature maps to the input value (latent space) of the DCGAN model, which is an adversarial image generation model using a convolutional neural network. To solve the problem that the neural network model cannot generate clear images in case of lack of training data, one additional feature map was added to the input value of the generation model, latent space. The feature map was extracted from 2,000 images of the CelebA dataset consisting of human face images through principal component analysis. We used 3,838 Large-Age-Gap datasets and one feature image for training. Compare to the previous model which uses 200,000 images, the proposed model generates more natural facial images with only 3,829 examples and the error rate is significantly reduced than the previous model at the beginning of the model training.