Adaptive feature refinement for texture-preserving single image super-resolution

Abstract Single-image super-resolution (SISR) has become a major focus in the field of computer vision, with significant applications in industries such as medical imaging, satellite analysis, and security surveillance. Recent developments have led to the use of deep convolutional networks and generative adversarial models, such as ESRGAN, which applies residual-dense connections to reconstruct high-resolution (HR) images from low-resolution (LR) inputs. Nevertheless, these architectures often fail to capture long-range dependencies and the most delicate textures that are essential for photo-realistic restoration. In the present work, we propose a modified ESRGAN model by integrating a Convolutional Block Attention Module (CBAM) into the Residual-in-Residual Dense Block (RRDB) structure and replacing the final dense layer with a more advanced feature recalibration module. This modification introduces a slight computational overhead but substantially enhances attention-driven texture refinement. Experiments conducted on the Div2K, BSD100, and Set14 datasets demonstrate that the CBAM-ESRGAN model outperforms existing state-of-the-art techniques, achieving superior PSNR, SSIM, LPIPS, and Perceptual Index scores, while also improving visual quality and reducing both inference time and model complexity. Additional experiments and their corresponding analysis further clarify the optimal placement of the CBAM module, considering the trade-off between performance and computational efficiency. The proposed model is intended for implementation as a practical alternative to existing high-quality super-resolution methods in both real-time and resource-constrained environments.