ECSANet: Enhancement Cross-Aggregation Statistical Attention Network for Image Style Transfer

Attention-based style transfer methods have made remarkable advancements in generating high-quality stylized images. However, they struggle to produce stylized images with consistent style distribution with real-style images while preserving the content structure. To this end, we propose a novel Enhancement Cross-Aggregation Statistical Attention Network (ECSANet) for image style transfer. Specifically, we design an enhancement cross-aggregation statistical attention (ECSA) module. It achieve can refine content feature and enhance style semantics by learning channel relevance and global dependency in addition to depict the captured style sematics into the content structure by a cross-aggregation statistical attention approach. Furthermore, we construct a statistical-aware skip connection (SSC) between the encoder-decoder. It can dynamically inject shallow information into the decoder, improving content retention and overall stylized quality. Extensive experiments have verified the effectiveness of our proposed ECSANet, surpassing previous state-of-the-art attention-based style transfer methods.