Learning to Correct: Efficient Image Super-Resolution via Fourier-Based Correction Maps

Image super-resolution (SR) remains a fundamental challenge in computer vision, particularly when deploying lightweight models in resource-constrained environments. While frequency-domain modeling offers a global receptive field with efficient complexity, traditional Fourier-based approaches often suffer from spectrum shifting and spatial artifacts. In this paper, we propose SpecResNet, a lightweight SR framework (965K parameters) centered on the paradigm of Correction Map Prediction. SpecResNet learns to predict a normalized correction map to restore high-frequency details missing from bicubic upsampled inputs. Our architecture incorporates an Fourier Block with learnable transformations and Dense Fourier Mixing to stabilize spectral activations and maximize feature reuse. Experimental results demonstrate that SpecResNet achieves state-of-the-art performance on benchmark datasets, particularly in structural preservation, while maintaining a compact 3.68MB footprint.