Blurred image restoration method based on coupled generative adversarial network

As high-resolution display technologies advance, obtaining sharp imagery during dynamic screen transitions becomes increasingly important. However, motion blur caused by high-speed imaging often obscures fine details such as dead pixels and flicker, thereby degrading visual fidelity. Meanwhile, classical deconvolution methods are computationally expensive and sensitive to noise, and deep learning approaches typically require paired training data that are difficult to acquire. To address these limitations, this paper presents an unsupervised motion-deblurring framework based on a coupled generative adversarial network. The model employs dual generators and dual discriminators to enable bidirectional translation between blurred and sharp domains without paired supervision. Each generator integrates an Inception-Res-Net backbone with a Feature Pyramid Network for multiscale feature extraction, together with a global residual connection to preserve structural integrity. Training is guided by a composite objective that combines adversarial, perceptual, and reconstruction losses, promoting stability and high-quality restoration. Experiments on the GoPro dataset demonstrate that the proposed method surpasses conventional unsupervised baselines in both PSNR and SSIM, and qualitative results confirm superior recovery of fine details. By learning from unpaired data, the method exhibits strong generalization and provides a practical solution for high-fidelity motion deblurring.