Image Deblurring Using Fusion Transformer-Based Generative Adversarial Networks

Using the Transformer for motion deblurring enables a broader receptive field, and by stacking multiple Transformer modules, it captures global correlations in features. However, this increases network complexity and poses convergence challenges. To address this, a Generative Adversarial Network called XT-GAN, which combines multiple-scale Transformers, has been proposed.XT-GAN leverages pyramid features from a convolutional network as a lightweight substitute for multi-scale inputs. Within the output pyramid convolutional features, different-scale features are computed in parallel using multi-head self-attention. These features are combined with a proposed feature enhancement module to represent information at different scales. Finally, the network outputs from various modules are concatenated and restored to the original image size.In experiments conducted on the synthetic dataset GoPro, XT-GAN outperformed ordinary networks such as DeblurGAN, DeepDeblur, and SRN. It achieved a reduction in computational complexity of at least 70% while achieving PSNR and SSIM values of 29.13dB and 0.923, respectively. XT-GAN also demonstrated good robustness in the real dataset RealBlur-J, with PSNR and SSIM values of 28.40 and 0.852. It effectively handles motion blur in real-world scenarios, suppresses image artifacts, and restores natural and clear details.