Anemone Image Generation Based on Diffusion-Stylegan2

Abstract

Given the complexity and uncertainty of the underwater environment, it is of great importance to generate realistic and high-quality images. In this paper, we propose six unconditional generative models based on the Diffusion-Styegan2 generative model, incorporating Wasserstein, R2 regularization terms, and other techniques for anemone image generation. The Wasserstein distance technique is used in the loss part of Diffusion-Styegan2, combined with the back propagation algorithm to compute the gradient in the neural network while retaining the computational map to improve the training efficiency and training stability; the R2 regularization term is used to introduce the r2 hyperparameter, and the L2 regularization technique is used based on the original R1 regularization term to regularize the gradient of the discriminator to improve the training and generation performance of the model; the ADA technique is used based on DWBG-Stylegan2 to further improve the quality and stability of the generated images. In addition, a set of SA datasets (sea anemone datasets) with a resolution of $256^{\ast} 256$ is proposed in this paper. The experimental results show that the FID value of Diffusuion-Stylegan2 is 10.31, the value of FID of DWBG-Stylegan2 is 8.32, the value of FID of Diffusion-Stylegan2-R2 is 9.58, and the optimal FID value of this experiment is achieved by DWBG-Stylegan2-ADA with a value of 5.67, which is considerably lower compared to the FID value of Diffusion-Stylegan2. Therefore, techniques such as Wasserstein and R2 regularization terms can effectively generate more realistic images of anemones. Meanwhile, this experiment provides new ideas and methods for the construction of the unconditional generative model.

Keywords

• Diffusion-Stylegan2
• Wasserstein
• R2 regularization term
• SA datasets