Lightweight Implicit Blur Kernel Estimation Network for Blind Image Super-Resolution

Blind image super-resolution (Blind-SR) is the process of leveraging a low-resolution (LR) image, with unknown degradation, to generate its high-resolution (HR) version. Most existing blind SR techniques use degradation estimator network explicitly estimate blur kernel guide supervision ground truth (GT) kernels. To solve this issue, it necessary design an implicit that can extract discriminative representation without relying on ground-truth We lightweight approach for estimates and restores HR based deep convolutional neural (CNN) residual generative adversarial network. Since unknown, following formation model problem, we firstly introduce network-based kernel. This achieved by (i) Super Resolver that, from input, generates corresponding image; (ii) Estimator Network generating input datum. The output both models used in novel loss formulation. proposed end-to-end trainable. methodology substantiated quantitative qualitative experiments. Results benchmarks demonstrate our computationally efficient (12x fewer parameters than state-of-the-art models) performs favorably respect approaches be devices limited computational capabilities.