CBAM-Enhanced CycleGAN for Preserving Cultural Symbols in Artistic Design

Cultural and creative design (CCD) requires a balance between artistic style transfer and cultural symbol fidelity. Nevertheless, general style transfer models generally exhibit drawbacks when applied to CCD. Typical limitations include insufficient attention to key cultural regions, the absence of multi-scale structural constraint mechanisms, and inadequate fusion of shallow texture details and deep semantic features. These limitations lead to problems such as structural distortion, loss of decorative patterns, and reduced recognizability of cultural symbols in generated outputs. In order to handle the above challenges, this work proposes an improved Cycle-Generative Adversarial Network (CycleGAN) architecture, integrating three targeted modules. The convolutional block attention module (CBAM) strengthens focus on key cultural regions; the multi-scale discriminator reinforces constraints on global and local style consistency; the shallow and deep feature fusion mechanism better preserves texture details and semantic integrity. Objective indicator evaluation and subjective human assessment are conducted on the WikiArt dataset, cultural heritage visual dataset, and multiclass visual arts dataset. Experimental results show that the model achieves a style accuracy of 96.53%, a peak signal-to-noise ratio of 22.43 decibels (dB), and a structural similarity index of 0.792. Cultural feature fidelity tests reveal that the detection confidence of cultural symbols decreases by only 8.5%, and the average intersection over union reaches 0.91. Subjective evaluation further verifies the model’s superiority in stylistic realism and the integrity of cultural symbols. The above outcomes demonstrate that the proposed model can provide robust and practical design assistance tools for the generation of intelligent cultural and creative products. At the same time, it has direct application value for digital heritage conservation and creative industries.