Image Fidelity in Generative Adversarial Network: Challenges and Solutions for Small Datasets

University of International Business and Economics, No.10 Huixin Dongjie, Chaoyang District, Beijing, China

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

This paper systematically summarizes the challenges and countermeasures in image generation when the training samples of the generated countermeasure network (GAN) are very few (less than 100 – 1000). The research proposes a cross method comparison framework, which classifies the existing strategies into three categories: data enhancement (such as DiffAug and ADA), regularization and structure improvement (such as ContraGAN and FastGAN), and cross domain adaptation (such as FreezeD and SSR). Experiments show that DiffAug performs better under very low sample size, and ADA is suitable for slightly large-scale data; Regular methods improve diversity and structural integrity, and migration strategies enhance cross domain performance. The study emphasized that the generation quality should be evaluated by combining multiple indicators such as KID and precision recall rate, and pointed out that in the future, attention should be paid to enhanced regular joint scheduling, generalized migration paradigm and more reliable evaluation system.