Face Image Generation with Features of Specific Group from Small Dataset

Abstract

The research on face image generation has gained widespread attention in the field of image generation. However, large datasets are required to train a generative model to produce images of high quality and resolution. It is especially difficult to collect multiple face images of a specific group. It may also be difficult to establish an efficient model by training with a small dataset.This paper presents a novel method for the generation of face images with the features of a specific group. The face images of a dataset are embedded into a latent space as sets of latent variables by Image2StyleGAN and are expressed as a distribution of the sets of latent variables in the latent space of a pre-trained StyleGAN2. Principal Component Analysis (PCA) is used to extract the features of the distribution. The generation of images for different groups does not require re-training of the model, as a pre-trained model is used instead. Furthermore, the proposed method can generate high-quality images with the features of the group from a dataset with only about 100 face images. However, the quality and the variety of the generated images can vary depending on a Cumulative Contribution Rate (CCR) of the PCA. Therefore, this study also proposes a metric called the Fréchet Inception Distance in Principal Component Space (FID-PCS), which can evaluate the generated images even with a small dataset. The FID-PCS can be used to determine the CCR that generates images with a good balance between the quality and the variety. The face images of three groups were collected as datasets to evaluate the validity of the proposed method, which include male idols, female idols, and male mixed martial artists. It was observed that images with the features of the group are generated by the face distributions extracted by the PCA, and the images with high quality and wide variety are generated by determining the appropriate CCR by the FID-PCS.