An enhanced CycleGAN approach for landscape design: Style transfer and color harmonization

Style transfer and color optimization are critical aspects of landscape design, directly influencing the aesthetic and functional quality of design outputs. However, existing GAN-based methods struggle with content preservation, color consistency, and computational efficiency in complex scenes. This paper presents a novel CycleGAN model specifically designed for landscape design, leveraging ResNeXt for hierarchical feature extraction to capture intricate landscape details, and SE attention to emphasize key elements such as foliage and terrain. Depthwise separable convolutions reduce inference time by 40 percent, making the model more efficient. A refined loss function, combining adversarial, cycle consistency, color consistency, and contrast losses, ensures both style consistency and content fidelity. Experimental results show the model outperforms state-of-the-art methods on multiple datasets, achieving significant improvements in both quantitative metrics and visual quality. It reduces FID by 21.3 percent and offers real-time performance with 31.27 ms per image on the Summer2Winter dataset, effectively balancing style transfer and content preservation for a wide range of landscape design tasks. This work provides an efficient solution that contributes to the advancement of intelligent and innovative landscape design techniques.