GAN Evaluation by Multi-Method Fusion

Abstract

Generative adversarial networks (GAN) have made great progress in areas, such as computer vision, but evaluation of GAN models is still a daunting task. Numerous researchers have proposed many methods for GAN model evaluation. However, most methods have strong pertinence, which leads to large deviations in the results of different evaluation methods. To date, no consensus has been reached on which method best captures the advantages and limitations of GAN models. In this study, a multi-method fusion approach is proposed to evaluate GAN models from many aspects, such as accuracy, diversity of the images generated by the model, and similarity with the training images. This method provides a new idea for GAN model assessment. Meanwhile, a strategy similar to ensemble learning is used to analyze the generalization of multi-method fusion for GAN assessment. This method and the human eye cognitive assessment model are used on celebA benchmarks and self-acquired image data sets to generate results. Through human eye cognition and comparison with the latest GAN model assessment method, showing that the multi-method fusion GAN evaluation method proposed in this paper has strong robustness and effectiveness.