Computer and Modernization ›› 2024, Vol. 0 ›› Issue (06): 14-18.doi: 10.3969/j.issn.1006-2475.2024.06.003

Previous Articles     Next Articles

Semi-supervised Image Generation Model Based on StyleGAN

  



  1. (Department of  Information, Beijing University of Technology, Beijing 100020, China)
  • Online:2024-06-30 Published:2024-07-17

Abstract: Abstract: This paper introduces SG-GAN, a semi-supervised StyleGAN model that overcomes the limitations of traditional StyleGAN. The quality of generated images using StyleGAN is heavily dependent on the quality of the training data set. When the training image quality is low, StyleGAN often fails to generate high-quality images. To address this issue, SG-GAN generates and trains support vector machine(SVM)training samples based on the one-to-one correspondence between vectors w and images in StyleGAN. SVM and StyleGAN mapping network are then used to screen vectors w before generating each image to improve the quality of the resulting images. For batch image generation, gene vectors are generated by the gene vector generator and combined randomly while all permutations of style vectors are obtained using a dynamic cycle backtracking algorithm. Individuals are generated from the permutation results and screened for excellence using an evaluation function after multiple iterations. Experiments were carried out on open data sets and compared with other advanced methods, demonstrating that SG-GAN improves upon StyleGAN's accuracy significantly. The model achieves FID 2.74, an accuracy rate of 74.2%, and a recall rate of 51.2% on the lsun cat face data set, further validating the efficacy of the model. At the same time, the model achieved an accuracy of over 70% on the Cat Dataset, CIFAR-100, and ImageNet datasets, thereby verifying its good generalization ability.

Key words: Key words: generative adversarial network, genetic algorithm, style vector, support vector machines, dynamic loop backtracking

CLC Number: