Image Super-Resolution Based on Fuzzy Kernel Singular Value Decomposition Networks

Super-resolution (SR) technology aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs. To address the issue that traditional image superresolution methods tend to suffer from detail loss when dealing with complex degradation scenarios, this paper proposes an optimized framework integrating BERT and Generative Adversarial Networks (GANs). Guided by text prompts, this framework extracts accurate semantic and degradation features via the BERT encoder, which synchronously facilitates the collaborative training of the generator and the discriminator. Meanwhile, it interfaces with the blur kernel calculation, Schur complement, and post-processing modules, and further optimizes the reconstruction quality by incorporating the Zero-Shot Super-Resolution (ZSSR) technology. Experimental results demonstrate that the proposed method achieves superior performance on multiple datasets. Compared with state-of-the-art methods, it yields an improvement of approximately 2 dB in key metrics such as PSNR, and can efficiently generate high-fidelity HR images with rich details, providing a novel and effective solution for image super-resolution tasks in practical scenarios.