Optimizing Generative Adversarial Networks Models for Non-Pneumatic Tire Design: A Comparative Analysis and Evaluation

Optimizing Generative Adversarial Networks Models for Non-Pneumatic Tire Design: A Comparative Analysis and Evaluation

연구 분야: Artificial Intelligence

학회: RACS '23: Proceedings of the 2023 International Conference on Research in Adaptive and Convergent Systems

Pneumatic tires1 are used in a wide range of industries. However, they are difficult to design and rely on the knowledge of experienced designers. To aid in the design of pneumatic tires, this paper suggests the use of Generative Adversarial Networks (GAN) models. The 2000 created images for training were used after removing highly sim-ilar images using Mean Squared Error (MSE) and Structural Similarity Index (SSIM). To find the best model for generating patterns of regularly shaped non-pneumatic tires, GAN, Deep Convolu-tional Generative Adversarial Networks (DCGAN), StyleGANv2 ADA and ProjectedGAN were compared and analyzed. In the qualitative evaluation, the GAN, DCGAN, and StyleGAN v2-ADA models showed that the circle shape was distorted and did not maintain a consistent pattern, but ProjectedGAN showed that the circle remained consistent and the pattern was less distorted than the other GAN models. When evaluating quantitative metrics, ProjectedGAN performed best among several techniques that measure the difference be-tween the generated and actual image distributions.