Art Painting Style Transfer Algorithm Based on Deep Learning and Visual Attention Mechanism
C. T. Dai, X. Zhou · Advanced Electromagnetics · 2026
Robust semantic feature preservation and adaptive multi-scale information fusion are critical challenges in intelligent visual information processing and communication-enabled perception systems. To address structural distortion and semantic inconsistency in complex image style transfer, this study proposes a dual visual attention framework that integrates spatial attention and channel attention for adaptive feature enhancement and style-aware representation learning. A pre-trained VGG-19 network is employed to extract hierarchical semantic features, while the spatial attention module strengthens foreground structural preservation through semantic guidance and the channel attention module dynamically selects discriminative style representations using global feature responses. The weighted features are collaboratively fused through adaptive instance normalization and reconstructed by an end-to-end decoder under the joint optimization of content loss, style loss, and region consistency loss. Experimental results demonstrate that the proposed method achieves superior structural similarity, semantic preservation, and style fidelity compared with representative baseline algorithms, maintaining an SSIM above 0.85 under complex scenarios while significantly reducing style distortion and improving image quality. Beyond artistic image synthesis, the proposed architecture provides an effective framework for semantic information fusion, adaptive feature transmission, intelligent visual perception, and multi-scale representation learning, offering valuable methodological references for communication-oriented sensing systems, image-guided information processing, and engineering applications related to Electromagnetic Waves, Antennas and Propagation.