A Sensitive Image Generation Method Based on Improved PatchGAN
Shilin Zhang, Hongliang Wang, Liang Wang · 2023
Aiming at the problem of less data and unbalanced number of sensitive image samples in the deep learning task of sensitive image detection, this paper proposes a sensitive image generation method SI-GAN based on improved PatchGAN. The SI-GAN generator uses the DUD module, and uses the concat method to increase the number of channels of the feature map in the quantitative dimension, which improves the generation ability of the generator. The SI-GAN discriminator uses the RBL module to improve PatchGAN, extract more advanced feature information, and increase the discriminant's discriminative ability. The experimental results show that SI-GAN has better results in terms of generated image quality. The FID value of SI-GAN is 5.87 lower than that of PatchGAN. The IS value of SI-GAN is 5.09 more than PatchGAN. The SI-GAN proposed in this paper is able to optimize the quality of the generated images and generate images with more details.