A conditional generative adversarial network with dual-context aggregation and color enhancement for digital mural creation

Abstract This paper proposes an improved model based on generative adversarial networks to address the problems of unreasonable structural layout, color gradient distortion, and limited style controllability in digital mural generation. This model integrates the Dual Aggregation Context Module, Fast Fourier Convolutional Global Encode, Color Enhancement Polarized Self Attention Mechanism, and Pre trained Style Encoder, aiming to enhance the structural integrity, color authenticity, and style diversity of mural generation through multi-scale feature fusion, frequency-domain spatial domain joint modeling, and color space optimization. Experiments on Dunhuang murals and MetCollection datasets have shown that this method significantly outperforms mainstream generative models in objective evaluation metrics. The introduced evaluation of interaction efficiency and style consistency shows that this method can achieve higher consistency with the target style (CLIP similarity 0.891) with fewer interaction rounds (average 2.3 rounds) and faster response time (153 ms). In subjective evaluation, the average scores of generated images in terms of aesthetics, style consistency, and other dimensions are all higher than 4.3 points. The results indicate that the proposed method can effectively enhance the quality control of the structure, color, and style of digital murals, providing a reliable technical path for the digital restoration and creative generation of cultural heritage.