GAN-Based Blurred Image Recovery for Moving Targets

Image restoration is crucial for image quality, yet existing methods suffer from high cost, distortion, and poor generalization. This paper proposes a GAN-based model for recovering motion-blurred images with improved coherence and texture. We introduce hierarchical noise injection, which splits and injects noise across multiple layers with class labels to enhance multi-level feature learning, and a dual-branch decoder that integrates global and spatial information for realistic textures. To reduce GAN instability and mode collapse, we employ a coupled discriminator with feature matching loss and a progressive reconstruction strategy. Experiments on GoPro, REDS, and DVD datasets, with comparisons to BigGAN and RFR, show our method achieves superior restoration performance and robustness, validated through comparative and ablation studies.