Deep Learning Based Efficient Single Image Super Resolution

Blind image super-resolution (Blind-SR) involves recovering a high-resolution (HR) image from its low-resolution (LR) counterpart under unknown degradation conditions. Existing approaches often rely on explicit degradation estimators that require ground-truth information about the degradation kernel, which is challenging to obtain in real-world scenarios. Implicit degradation estimators offer an alternative but typically suffer from a performance gap compared to explicit methods, particularly in computational efficiency and accuracy. In our first study, we addressed these challenges by designing a lightweight end-to-end framework for Blind-SR. This method integrates a deep convolutional neural network (CNN)-based Estimator module to implicitly estimate the blur kernel and a super-resolution residual convolutional generative adversarial network (Super Resolver) to reconstruct the HR image. The proposed model employs a novel loss formulation and achieves competitive performance on benchmark datasets, with a computational efficiency advantage—12× fewer parameters compared to state-of-the-art methods—making it suitable for devices with limited computational capacity. Building on this foundation, our second study introduced an enhanced approach to implicit blind-SR by developing a novel loss component that allows the implicit learning of degradation kernels without ground-truth supervision. We also designed a learnable Wiener filter module that efficiently performs deconvolution in the Fourier domain via a closed-form solution and a transformer-based refinement module to reconstruct the final HR image. Our model IDENet achieved significant performance improvements, outperforming existing implicit methods by 3dB PSNR and 8.5% SSIM on average while narrowing the gap with explicit methods to only 0.6dB PSNR and 0.5% SSIM. Remarkably, these results were obtained with 33% and 71% fewer parameters than state-of-the-art implicit and explicit methods, respectively. In our final study, we further refined the implicit blind-SR framework by introducing a degradation-conditioned prompt-learning module. This module leverages the estimated kernel to focus on discriminative contextual features, improving the reconstruction process. Our model, named PL-IDENet, demonstrated significant gains over state-of-the-art methods, achieving more than 0.4dB and 1.3% PSNR and SSIM improvements over the best implicit methods and 1.4dB and 4.8% over the best explicit methods. These results were achieved while maintaining a significantly lower computational complexity, with 25% and 68% fewer parameters than the best implicit and explicit methods, respectively. Together, these studies contribute to the field of blind image super-resolution by offering lightweight, effective, and scalable solutions that bridge the performance gap between implicit and explicit degradation estimators, making them practical for real-world deployment.