Mixed self-attention–enhanced generative adversarial network for spatially variant blurred image restoration
Lijian Xiao, Wende Dong, Gang Luo, Muhan Ji, Geng Li, Zhenzhen Zheng · Journal of Electronic Imaging · 2025
Transformer-based deep neural networks have demonstrated impressive performance in low-level visual tasks such as image deblurring. However, these networks often struggle with restoring spatially variant blurred images, resulting in outputs that lack detailed textures. This indicates that current transformer-based models fail to fully exploit their potential. To address this issue, we propose a mixed self-attention block by introducing a global attention modulator into the window–based self-attention mechanism. This approach leverages the complementary strengths of local and global feature extraction. We integrate this block into a U-shaped network, serving as the generator in a conditional generative adversarial network model. We further introduce a comprehensive loss function composed of adversarial loss, content loss, and edge loss for network training. Restoration quality is evaluated using both pixel-level and perceptual metrics to comprehensively assess visual fidelity. Moreover, the proposed model is specifically designed to handle spatially variant blur by combining localized window attention with global modulation across windows. This enables our model to adaptively focus on regions with varying blur severity and significantly improves restoration robustness under spatially variant conditions. Extensive experiments confirm the method’s superior ability to restore spatially variant blurred images, achieving competitive performance on challenging real-world datasets.