Multi-Discriminator Driven Enhanced Super-Resolution GAN: Combining Feature Clustering and Region-Based Optimization

To address the noise and artifact issues caused by single-discriminator architectures in complex image super-resolution tasks, this paper proposes the Multi-Discriminator Enhanced Super-Resolution GAN (MD-ESRGAN). The model constructs highly discriminative feature subsets through a feature clustering module (ISKM), combined with an image tile splitter (ITS) for localized detail enhancement, and dynamically routes features to dedicated expert sub-networks based on a feature routing mechanism (FRM). The innovative design of Feature Density Loss (LFD) optimizes the compactness of feature space distribution, while the Expert Load Consistency Loss (LELC) balances computational resource allocation. Experimental results show that MD-ESRGAN outperforms SwinIR on multiple datasets, significantly improving the quality of high-frequency detail reconstruction and structural fidelity, while exhibiting superior texture restoration ability in complex scenarios.