Abstract
The rapid advancement of AI technology has spurred evolution in deep learning applications such as image recoloring and creative style transfer. This paper develops an algorithm for rapid visual art style migration based on Cycle Generative Adversarial Networks (CycleGAN). Using the CelebA-HQ celebrity faces dataset and product design images, the system employs a preprocessing stage involving image resizing and Min-max normalization to ensure stable GAN training. A feature extraction technique, Histogram of Oriented Gradients (HOG), is utilized to evaluate structural consistency before and after migration. The CycleGAN produced high reconstruction quality with a PSNR of 23.84 and strong structural consistency with an SSIM of 0.81, confirming its effectiveness in artistic style migration for packaging design. To enhance the effect, images are identified using a multiscale discriminator, making the strategy highly applicable for commercial advertising.
Introduction
In industrial and visual communication design, integrating complex artistic styles into product packaging is traditionally a labor-intensive manual process. While style transfer is popular in general computer vision, applying it to commercial products presents two major hurdles: maintaining structural integrity and overcoming data scarcity. Brand identity relies on the "bones" of a product—its shape and logo—remaining recognizable. Furthermore, high-performing GANs often require paired datasets that are rarely available in design workflows. This research addresses these challenges by refining the CycleGAN architecture to handle unpaired data while preserving structural coherency.
Methodology
The core of the proposed system is a CycleGAN framework, which utilizes cycle consistency loss to translate images between domains without paired examples. To enhance performance for product design, two technical improvements were introduced:
- Multiscale Discriminator: Multiple discriminators operate at different scales to capture both fine-grained textures (brushstrokes) and global structural layouts simultaneously.
- HOG-based Feature Extraction: Histogram of Oriented Gradients (HOG) evaluates structural consistency by monitoring intensity gradients. This ensures the essential shape of the packaging remains intact during texture modification.
Experimental Results
The model was benchmarked using the CelebA-HQ dataset (30,000 images) and a custom Art Style Migration Performance Dataset. Performance was compared against Expo-GAN using Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity Index (SSIM).
The proposed CycleGAN achieved a PSNR of 23.84, significantly higher than Expo-GAN’s 21.39, indicating superior reconstruction fidelity and reduced noise. More importantly, the SSIM reached 0.81, confirming that the HOG-based approach successfully maintained the structural integrity of the original designs compared to the baseline.
Conclusion
This research demonstrates that CycleGAN, augmented with multiscale discriminators and HOG-based structural checks, is a powerful tool for automating artistic style transfer in product design. While the model excels at maintaining brand-critical geometry, future work will address challenges posed by highly abstract styles and high-density text layouts through adaptive attention mechanisms and domain-aware training strategies.
Summary updated