Advancements in High-Resolution Style Transfer: Unveiling the Precision-Enhanced-Cycle-GAN
Diye Xin · 2024
As Cycle-Consistent Generative Adversarial Networks enables style transfer between two distinct domains using unpaired data, allowing affective image-to-image transformation in scenarios lacking direct corresponding datasets. Sometimes, the low accuracy of the generated images, the requirement of the real-time processing and the single style of synthesis are questions. We demonstrate a new approach which can raise the accuracy of the Cycle-GANs called Precision-Enhanced-Cycle-GAN (PEC-GAN), intends to optimize the process of the style transference while preserving the content integrity of the original image. The new network structure introduces an image matching mechanism at the output layer, using an evolution function to detect some low-quality generated samples and leave high-quality samples for the next training step. Additionally, we introduce an encoding mechanism to make artificial images with various kinds of style. Through improved network architecture and training strategies, our method demonstrates advantages in processing high-resolution images, while also being able to adapt to a wider range of artistic styles. The experimental results show that this method can obtain target images faster than traditional style transfer techniques while also maintaining real-time processing capabilities, demonstrating technological progress and application potential.