Spectral Normalization for Generative Adversarial Networks for Artistic Image Transformation

Abstract

Artistic image transformation is a computer technique widely applied in art creation, design, entertainment, and cultural heritage by converting images into artistic styles. It offers innovative ways for artists to express themselves, provides designers with more choices and inspiration, enhances visual esthetics, and enables creative implementations in movies, games, and virtual reality. Additionally, it aids in the restoration and preservation of ancient artworks, allowing a deeper appreciation of classical art. Traditional image transformation methods, though effective for simple effects, lack the flexibility and expressiveness of deep learning–based approaches. To enhance the effectiveness and efficiency of artistic image transformation, this paper employs generative adversarial networks (GANs), which utilize an adversarial training mechanism between a generator and a discriminator to produce high-quality and realistic image transformations. This study introduces spectral normalization (SNGAN) to further improve GAN performance by constraining the spectral norm of the discriminator’s weight matrix, preventing gradient issues during training, thus improving convergence and image quality. Experimental results on the CHAOS dataset indicate that the proposed SNGAN model achieves the lowest mean absolute error (MAE) of 0.3420, the highest peak signal-to-noise ratio (PSNR) of 32.1423, and a structural similarity index (SSIM) of 0.6696, closely matching the best result. Additionally, the SNGAN model demonstrates the shortest training time, highlighting its efficiency. These results confirm that the proposed method achieves more realistic and efficient artistic image transformations compared to traditional methods and other deep learning algorithms.