Visual effects of urban environment art design based on composite Cycle Generative Adversarial Network

Abstract

In recent years, many regions have invested significant efforts in the in-depth study of artistic visual effects in urban environments to enhance the aesthetic appeal of urban districts. However, traditional visual optimization models are unable to refine image edges and textures, leading to distorted results. To optimize the artistic effects of urban district images, this study proposes a Cycle Generative Adversarial Network that incorporates a multi-head self-attention mechanism to enhance design. The network architecture is modified by adding multi-head self-attention modules. Spectral normalization is also introduced to improve the optimization process of both the generator and the discriminator. Finally, the images generated at different layers are concatenated, and the average of the extracted features is used as the output. The study compares the proposed network with publicly available datasets and commonly used optimization networks. The generated images achieve a pixel accuracy of 96.8% and a classification accuracy of 97.4%. In validation tests with a self-built dataset, the Peak Signal-to-Noise Ratio reaches 34.2 dB, outperforming the comparison models. In a comprehensive evaluation across various metrics, the Structural Similarity Index is 98.23% and the Visual Semantic Similarity is 96.78%. These results demonstrate that the proposed model analyzes images in various artistic styles and optimizes the visual effects of urban environments. It is beneficial for creating urban districts with distinctive artistic styles, enhancing the local cultural atmosphere, and attracting more domestic and international tourists.