Research Article
A Novel AI-Based Visual Stimuli Generation Approach for Environment Concept Design
Table 1
The characteristics, corresponding advantages, and training parameters of the six models.
| | Model | Feature of model | Corresponding advantages | Training parameter | | Epoch | Learning rate | Optimizer | Total parameter (M) |
| | DCGAN | Effective combination of GAN and CNN | (1) Improvement of the depth of the feature extraction network; (2) enhancement of image resolution | 8 × 103 | 2 × 10−3 | Adam | 12.33 | | WGAN | Wasserstein-distance-based GAN | (1) Improvement of training stability; (2) improvement of generated image quality | 8 × 103 | 2 × 10−4 | Adam | 23.36 | | LSGAN | Replacement from cross-entropy loss to least squares loss | (1) Mitigation of gradient disappearance problem; (2) improvement of generated image quality | 8 × 103 | 2 × 10−4 | Adam | 23.36 | | WGAN_gp | Introduction of the gradient penalty | Improvement of training stability | 8 × 103 | 2 × 10−4 | Adam | 23.36 | | DRAGAN | Introduction of a novel gradient penalty scheme | Improvement of stability | 8 × 103 | 2 × 10−4 | Adam | 23.36 | | PGGAN | Introduction of a novel progressive GAN training model | Solution of high loss problem for generating high-resolution images | 8 × 105 | 2 × 10−3 | Adam | 46.14 |
|
|
