Blind image super-resolution (SR) aims to recover high-resolution images without knowing the degradation. Existing methods often require ground-truth kernels and heavy networks. This work proposes a lightweight model (PL-IDENet) that implicitly learns degradations using a novel loss and a learnable Wiener filter. The method achieves better accuracy with significantly fewer parameters and computations.

Blind image super-resolution (SR) restores high-resolution images from low-resolution inputs with unknown degradations. Existing methods need ground-truth degradation or are computationally heavy. The proposed model uses a novel loss and a learnable Wiener filter to implicitly estimate degradation and efficiently solve deconvolution. It outperforms implicit SR methods and matches explicit ones with much fewer parameters..

The NTIRE 2024 Image Super-Resolution (×4) Challenge focused on enhancing low-resolution images using bicubic downsampling inputs. With no limits on model size or training data, the competition aimed for top PSNR performance on the DIV2K test set. Attracting 199 registrants and 20 final submissions, the challenge advanced state-of-the-art SR techniques and showcased emerging trends.

Blind image super-resolution (Blind-SR) restores high-resolution images from low-resolution inputs with unknown degradations. Existing methods rely on ground-truth blur kernels, but this work proposes a lightweight, implicit kernel estimation network (LBKENet) that learns without ground-truth supervision. It combines a super-resolver and a blur kernel estimator in an end-to-end framework with a novel loss design. The approach achieves competitive performance with 12× fewer parameters, making it suitable for low-resource devices.

We propose a lightweight blind super-resolution (Blind-SR) model that estimates blur kernels and restores HR images without ground-truth supervision. Our method uses a Super Resolver and an Estimator Network trained with a novel loss for joint kernel and image recovery. We further extend our work to handle anisotropic Gaussian kernels for more complex degradations. Experiments show our approach is efficient and performs well with significantly fewer parameters than state-of-the-art models.